Machine learning prediction model for treatment response

Abstract

The present invention is directed to a computer-implemented method for predicting a treatment outcome (patient treatment response or treatment efficacy) based on multi-modal features of a patient collected at at least two different time points. In particular, this invention relates to a method for predicting the response of lung cancer patients to immunotherapy.

Description

Technical Field

[0001] The present invention relates to the field of multimodal data integration and predictive models. In particular, the present invention relates to a computer-implemented method for predicting treatment response or treatment efficacy for a particular patient during the course of treatment. In particular, the present invention relates to a computer-implemented method for predicting treatment response or treatment efficacy over time for cancer patients such as lung cancer patients.

Background Art

[0002] For many diseases, there are one or more choices of treatment regimens that can be implemented, and that choice for a particular patient is usually based on the clinical status of those patients. Therefore, it is valuable to propose a predictive model for the clinical benefit associated with a particular treatment for a particular patient suffering from, for example, cancer (e.g., lung cancer, breast cancer or kidney cancer), a neurological disorder or a genetic disorder (e.g., a heart disease or a neurological disorder).

[0003] Triple-negative breast cancer (TNBC) is a biologically and clinically heterogeneous disease, which is associated with a poor prognosis compared to other subtypes of breast cancer. Neoadjuvant chemotherapy (NCT) is often performed prior to surgery, and achieving pathologic complete response (pCR) has been associated with improved long-term outcomes in terms of EFS (event-free survival) and OS (overall survival). Therefore, there is a high clinical interest in the ability to accurately predict data on the non-pCR state collected at baseline and during treatment planning or patient monitoring.

[0004] Surgical operation is the standard care for localized renal cancer. Imaging diagnosis plays a clinical role in staging and provides information about the scope of surgical resection (partial or radical nephrectomy, extended resection). In clinical routine, 15% of the tumors initially evaluated as T1-T2 by imaging diagnosis are upstaged to the postoperative pT3a state, which implies a high recurrence risk. The ability to correctly predict the pT3a state preoperatively can affect the appropriate surgical approach. Individually predicting the risk of upstaging of clinical T1 or T2 tumors to pT3a is, therefore, of high surgical interest.

[0005] Lung cancer has become a major public health burden. In 2020, lung cancer was estimated to be the cause of approximately 2.2 million new cancer diagnoses worldwide, and lung cancer-related deaths were approximately 1.8 million, constituting a major factor among cancer-related deaths (WHO, 2020). Among various types of lung cancer, non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancer cases. Among NSCLC, cases are further classified into adenocarcinoma, squamous cell carcinoma, and large cell carcinoma (Travis et al. 2015, J Thorac Oncol. 10(9):1243-1260).

[0006] First-line treatment of stage IV NSCLC requires systemic therapy. Historically, patients have been treated with combination chemotherapy regimens of two agents, such as gemcitabine, vinorelbine or taxane combined with platinum chemotherapy. Under such circumstances, current systemic therapy for NSCLC is selected according to the presence of specific biomarkers. Strong tumor driver mutations have been identified in genes such as EGFR, ALK, ROSI, BRAF and NTRK1 / 2 / 3 in subsets of NSCLC patients. Collectively, these mutations account for approximately 30% of NSCLC cases and qualify patients for specific biomarker-driven targeted therapies. Considering that patients eligible for such first-line targeted therapies currently represent only about 30% of all patients with stage IV NSCLC, the majority of NSCLC patients cannot benefit from this treatment approach (Arbour and Riely, 2019, JAMA. 322(8):764―774). For such patients, immunotherapy has dramatically changed the overall treatment landscape for stage IV NSCLC over the past few years.

[0007] Malignant tumors can overexpress PD-L1 as a mechanism of immune evasion, thereby downregulating the immune response against the tumor through the inhibitory effect of the PD-L1 / PD-1 interaction (programmed cell death (PD)-1 and anti-programmed cell death ligand 1 (PD-L1)) (Pardoll et al., 2012, Nat Rev Cancer. 12(4):252―264; Dong et al., 2002, Nat Med. 8(8)793―800). Antibodies against PD-1 or PD-L1 can block this interaction and bring about "release of the brake" during the anti-tumor immune response. This treatment strategy has been successful across many tumor types, including NSCLC (Pardoll et al, 2012, supra).

[0008] Under such circumstances, the immune checkpoint inhibitors pembrolizumab, nivolumab, and atezolimumab, which are anti-PD-(L)1 antibodies, have been approved by the FDA and EMA as monotherapy regimens in the second-line NSCLC treatment setting after progression with platinum-based chemotherapy (Herbst et al. 2016, Lancet. 387:1540 - 1550; Borghaei et al., 2015, NEJM, 373:1627 - 1639; Rittmeyer et al. 2017, Lancet 389:255 - 265). Such second-line studies have highlighted two important findings. First, although in some studies PD-L1 expression has been shown to have some predictive power for responses to immunotherapy using immune checkpoint inhibitors, this predictive value is not comparable to that of targeted therapies in patients with specific genomic driver mutations. Second, patients with lung adenocarcinoma harboring mutations in the EGFR gene or ALK gene showed a lower response to immune checkpoint inhibitors compared to wild-type tumors (Yang et al., 2020, Annu Rev Med, 71:117 - 136). In practice, patients with metastatic NSCLC who are suitable for targeted therapies and have no contraindications to immunotherapy are currently mainly receiving either pembrolizumab monotherapy when PD-L1 > 50% or a combination therapy of pembrolizumab + chemotherapy as first-line treatment. Subsequently, patients with PD-L1 > 50% may also receive a combination therapy of pembrolizumab + chemotherapy, but the practice is still evolving.

[0009] Despite such clinical expectations for immunotherapy, significant challenges remain. This is because the majority of NSCLC patients appear to have intrinsic resistance to immunotherapy and do not respond. Overall, only about 20 - 30% of patients treated with immunotherapy show an objective response, which varies depending on the immune checkpoint inhibitor and the clinical treatment setting. At the same time, patients are potentially exposed to serious side effects, particularly immune-mediated responses against healthy organs. In addition, immune checkpoint inhibitors are particularly expensive, with most treatment methods costing over $100,000 per patient per year, imposing a significant economic burden on the healthcare system.

[0010] Today, high-level PD-L1 expression of over 50% is the only standard predictive biomarker for the efficacy of immune checkpoint inhibitors as monotherapy in first-line NSCLC treatment settings (Remon et al., 2020, J Thorac Oncol 15(6):914 - 947). However, PD-L1 remains a suboptimal biomarker for immunotherapy response and has several factors that limit its clinical use. Differences in the test platforms, the use of various cut-off points for expression among different immunotherapy agents, and the heterogeneity of PD-L1 expression within the tumor are all critical points (Bodor et al., 2020, Cancer 126:260 - 270). Against this background, the predictive power of PD-L1 for immunotherapy response remains limited. In fact, NSCLC patients with tumor PD-L1 > 50% typically show only an ORR (objective response rate) of approximately 45%, and patients with tumor PD-L1 > 90% still only reach an ORR of approximately 60% (Aguilar et al., 2019, Ann Oncol. 30:1653 - 1659).

[0011] Some examples of systems and user interfaces have been proposed to predict the expected response of a specific patient population when undergoing a specific treatment (WO2020142551). However, these do not target the prediction of specific patient biomarker signs and provide insights at the population level rather than at the individual level.

[0012] In such a context, validating the effectiveness of an approach or biomarker that can predict the clinical effects associated with a specific treatment (e.g., treatment using immune checkpoint inhibitors (ICI)) is very important for the most effective, efficient, and cost-effective use of these therapies. This can enable the optimization of the allocation of financial resources in the healthcare system while proposing the most valuable treatment based on the patient's characteristics for a specific patient.

[0013] More generally, an approach that integrates data from various data modalities to predict a patient's response to treatment is being increasingly used (Baptista et al., 2021, Briefings in Bioinformatics, 22(1):360 - 379). The advantages of the approach of integrating multiple data modalities are being gradually established in numerous clinical applications such as predicting a patient's response to ICI treatment, predicting the complete response of neoadjuvant therapy, predicting the recurrence risk after surgery, or predicting overall survival. What is common in such applications is that the prediction of the response to treatment or treatment effect cannot be accurately made using a single biomarker because the response to treatment or treatment effect depends on numerous biological, clinical, and treatment-setting factors that can only be captured through the integration of multiple biomarkers frequently obtained using various data modalities. For example, US2021 / 090694A1 describes a framework for storing data from a number of data sources including clinical records, genomic information from tumor samples, and samples from normal patients, and describes making the framework accessible to assist in clinical decision-making.

[0014] The steps of collecting, storing, and displaying data are an integral part of the ways to facilitate their use for clinical decisions. However, understanding which biomarkers to use and how to integrate them into models is essential for the success of predicting patients' responses to treatments using a multimodal approach. The ways to demonstrate the possible differences and their prediction results are the differences in some of the multimodal approaches already mentioned for predicting ICI responses in patients and the accuracy of those predictions. For example, US2020 / 258223A1 describes a deep learning method for predicting the status of biomarkers from slides of pathological tissues.

[0015] Methods for integrating information from various data modalities for ICI response prediction have been shown to improve the accuracy of predicting patients' responses to treatments (Herbst et al., 2018, Nature 553, 446 - 454). Methods for integrating multiple data sources, that is, information from multiple data modalities, are generally referred to as multimodal methods. For ICI, the most accurate multimodal methods for ICI response prediction typically integrate information on the essential factors of tumors (TMB, PD-L1 expression) and exogenous factors (the micro-treatment settings of tumors, the functionality of the immune system), and also integrate indicators of tumor response to treatment (differences in ctDNA, radiomics) in the same way (Anagnostou et al., 2020, Nat Cancer 1, 99 - 111; Nabet et al., 2020, Cell 183, 363 - 376.e13; Cristescu et al., 2018, Science 362, eaar3593; Jiang et al., 2018, Nat Med 24, 1550 - 1558).

[0016] Implementing the most viable multimodal approach in clinical practice is difficult because the data (i.e., genomics and ctDNA profiling) that assist these predictions are costly and generally not obtained in clinical routine. To utilize the multimodal approach power to assist clinical judgment, including identifying patients most likely to respond to ICI treatment, it is important to increase the generality of the model and its applicability to clinical treatment settings.

[0017] Missing data points are a common problem in the development, testing, and implementation stages of multimodal prediction models, especially models developed for use in clinical treatment settings. Initial studies have shown that the level of data completeness depends on the type of data and can vary between 1.1% and 100% (Hogan and Wagner, 1997, J Am Med Inform Assoc 4(5):342-55). Missing data can be attributed to various sources, including missing collection or documentation, human error, processing errors, multifunctional devices, participants' refusal to answer questions, study termination, or inaccurate data merging. Regardless of its origin, missing data can induce systematic or random bias in data impact prediction performance. Despite being prevalent in clinical research, 152 recent surveys on machine learning-based clinical prediction models have shown that a significant portion of the models either ignore missing data or use deletion of records with missing data (complete case analysis) as the most common method to address this difficulty (Nijman et al., Feb 2022, JCE 142:218-229). Removal of incomplete cases can lead to the loss of a significant number of highly informative missing data. To avoid bias and loss of analytical power and prediction accuracy, devising alternative strategies for handling missing data is particularly important when using data from clinical routines where missing data are prevalent. A more appropriate approach for dealing with missing data is to implement a complementary model based on characteristics observed from other records / patients. In addition to imputation, several alternative methods are available (Gheyas and Smith, 2010, Neurocomputing 73(16-18):3039-3065), which include pattern mixture or surrogate branching and can also be used to avoid the need for imputation while limiting the bias induced by the removal of incomplete cases (Donders et al., 2006, JCE 59(10):1087-1091).However, which method is preferable remains unclear at present, and no consensus has been reached regarding the effectiveness of various methods and how they can be used depending on the extent or type of missing data (Nijman et al., supra). This is the case in the analysis of clinical records where missing data are prevalent (Beaulieu-Jones et al., 2018, JMIR Med Inform 6(1):e11).

[0018] Therefore, there is a general need to rapidly, efficiently, and comprehensively aggregate and analyze such multimodal patient data using data from approaches routinely used in clinics. In addition, such methods must be robust to the presence of missing data and similarly robust to replicated or redundant information to assist in clinical use.

Prior Art Documents

Patent Documents

[0019]

Patent Document 1

Patent Document 2

Patent Document 3

Non-Patent Documents

[0020]

Non-Patent Document 1

Non-Patent Document 2

Non-Patent Document 3

Non-Patent Document 4

Non-Patent Document 5

Non-Patent Document 6

Non-Patent Document 7

Non-Patent Document 8

Non-Patent Document 9

Non-Patent Document 10

Non-Patent Document 11

Non-Patent Document 12

Non-Patent Document 13

Non-Patent Document 14

Non-Patent Document 15

Non-Patent Document 16

Non-Patent Document 17

Non-Patent Document 18

Non-Patent Document 19

Non-Patent Document 20

Non-Patent Document 21

Non-Patent Document 22

Summary of the Invention

[0021] The present invention is based on the development of a specific computer-implemented method for predicting treatment response or treatment efficacy (i.e., generally treatment effect) for a specific patient based on at least two types of features of the patient selected from clinical features, biological features, genomic features, and radiological features, for example, at least clinical features and radiological features, preferably clinical features, biological features, genomic features, and radiological features (hereinafter referred to as multimodal features for at least two types of features in this specification). The computer-implemented method of the present invention uses the features of the patient that can be input into a machine learning algorithm, and the aforementioned features are pre-processed raw data of the patient selected from clinical data, biological data, genomic data, and radiological data. The computer-implemented method of the present invention uses a combination of a trained completion machine learning model trained to complement the missing features of the patient and a trained prediction machine learning model trained to predict the treatment response (or treatment efficacy) of the patient, and also uses a list of highly informative feature identifiers obtained during the training of this prediction machine learning model. It has been shown that the combination of multiple machine learning models leads to an improvement in the accuracy of this method.

[0022] The method of the present invention is particularly suitable for processing multimodal features or data received from real-world treatment settings, where usually some data is missing and the data contains a lot of noise. The method of the present invention is particularly suitable for using the clinical features or data of the patient, which are more easily accessible but, on the other hand, require a more comprehensive analysis as they are included in the aforementioned prediction.

[0023] The method of the present invention is particularly suitable for processing multimodal features derived from data received from real-world treatment settings, where the aforementioned features derived from the data are received at various time points. The method of the present invention makes it possible to explain the changes in such features derived from data over time, and thus provides additional input information that can be used in this analysis.

[0024] In one embodiment, a computer-implemented method for predicting a patient's treatment effect is provided, the method comprising the following steps: a) the following, i. a trained completion machine learning model trained to complement the missing features of a patient, ii. a trained prediction machine learning model trained to predict the treatment effect of a patient, and iii. a list of high-information-value feature identifiers used for training the prediction machine learning model, wherein the step of obtaining comprises using at least two types of features selected from among the clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be performed and undergoing the same treatment, and / or at least one longitudinal feature, to train the completion and prediction machine learning models and obtain the list of high-information-value feature identifiers, for each patient in the cohort, at least one of the multimodal features is collected at at least two time points, a metric for the change between the values of at least one multimodal feature of each patient collected at the at least two time points is calculated, and at least one longitudinal feature is obtained for each patient, and the obtaining step; b) receiving separately the multimodal features of the patient, the multimodal features of the patient comprising at least two types of features selected from among clinical features, biological features, genomic features, and radiological features, and the multimodal features of the patient being incomplete, wherein at least one of the multimodal features of the patient is collected at at least two time points, a metric for the change between the values of the at least one received multimodal feature of the patient collected at the at least two time points is calculated and at least one longitudinal feature is obtained, ​The step of receiving, which is performed before or after the step of complementing the multimodal features of a patient for whom the calculation of at least one longitudinal feature is missing, c) aggregating the multimodal features of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step; d) inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector; e) filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; f) inputting the predicted feature value vector into the trained prediction machine learning model and outputting a prediction of the treatment effect of the patient.

[0025] In another embodiment, a computer-implemented method according to the present invention is provided, which is performed before the step of complementing the multimodal features of a patient for whom the calculation of at least one longitudinal feature is missing, and the method includes the following steps: a) the following, i. a trained complementary machine learning model trained to complement missing features of a patient, ii. a trained prediction machine learning model trained to predict the treatment effect of a patient, and iii. a list of highly informative feature identifiers used for training the prediction machine learning model, the step of obtaining, using at least two types of features selected from among clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be performed and undergoing the same treatment, the complementary and prediction machine learning models are trained and the list of highly informative feature identifiers is obtained, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, a metric is calculated for the change between the values of at least one multimodal feature of each patient collected at the at least two time points, obtaining, for each patient, at least one longitudinal feature, the obtaining step; b) receiving separately the multimodal features of the patient, the multimodal features of the patient including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, and the multimodal features of the patient being incomplete, the receiving step, wherein at least one multimodal feature of the patient is collected at at least two time points; c) calculating a metric for the change between the values of at least one received multimodal feature of the patient collected at at least two time points to obtain at least one longitudinal feature of the patient; d) aggregating the multimodal features of the patient and at least one longitudinal feature of the patient into a feature value vector, the aggregating step, the feature value vector being incomplete; e) inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector; f) filtering the plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; g) inputting the predicted feature value vector into the trained predictive machine learning model to output a prediction of the treatment effect of the patient.

[0026] In another embodiment, a computer-implemented method according to the present invention is provided, The calculation of at least one longitudinal feature quantity is performed after the step of complementing the multi-modal feature quantity of the patient with missing data, and this method includes the following steps: a) The following i. A trained complementary machine learning model trained to complement the missing feature quantity of the patient, ii. A trained prediction machine learning model trained to predict the treatment effect of the patient, and iii. A list of high-information-value feature identifiers used for training the prediction machine learning model, which is a step of obtaining, using at least two types of feature quantities selected from the clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities of a cohort of patients having the same disease as the patient for whom the prediction is to be performed and undergoing the same treatment, and / or at least one longitudinal feature quantity, the complementary and prediction machine learning models are trained and the list of high-information-value feature identifiers has been obtained, For each patient in the cohort, at least one of the multi-modal feature quantities is collected at at least two time points, A metric for the change between the values of at least one multi-modal feature quantity of each patient collected at these at least two time points is calculated, For each patient, at least one longitudinal feature quantity is obtained, the obtaining step, b) A step of separately receiving the multi-modal feature quantity of the patient including at least two types of feature quantities selected from the clinical feature quantity, biological feature quantity, genomic feature quantity, and radiological feature quantity, the multi-modal feature quantity of the patient being incomplete, At least one multi-modal feature quantity of the patient is collected at at least two time points, the receiving step, c) A step of aggregating the multi-modal feature quantity of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step, d) Inputting the feature value vector into the trained completion machine learning model to output a complete feature value vector; e) Calculating a metric regarding the change between the values of at least one multi-modal feature of the received patient collected at at least two time points in the form of the complete feature value vector to obtain a complete longitudinal feature value vector; f) Aggregating the multi-modal feature of the patient in the complete feature value vector and at least one longitudinal feature of the patient in the complete longitudinal feature value vector to obtain a completely aggregated multi-modal longitudinal feature value vector; g) Filtering the multiple features of the completely aggregated multi-modal longitudinal feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector which is a subset of the completely aggregated multi-modal longitudinal feature value vector consisting of filtered feature values; h) Inputting the predicted feature value vector into the trained prediction machine learning model to output a prediction of the treatment effect of the patient.

[0027] In another embodiment, a computer-implemented method according to the present invention is provided, wherein the prediction of the treatment effect of the patient is represented as a prediction of the response of the patient to the treatment, and the trained prediction machine learning model is trained to predict the treatment effect of the patient represented as the response of the patient to the treatment.

[0028] In yet another further embodiment, a computer-implemented method according to the present invention is provided, wherein the prediction of the response of the patient to the treatment is classified as complete remission, partial remission, stable disease or progression, or as the probability of the response of the patient to the treatment.

[0029] In another embodiment, a computer-implemented method according to the present invention is provided, wherein the prediction of the treatment effect of the patient is represented as a prediction of the treatment efficacy of the patient, and the trained predictive machine learning model is trained to predict the treatment effect of the patient represented as the treatment efficacy of the patient defined as the length of time to an event.

[0030] In yet another further embodiment, a computer-implemented method according to the present invention is provided, wherein the treatment efficacy of the patient is defined as the length of time to an event and is selected from progression-free survival (PFS), overall survival (OS), duration of response (DoR), and time to progression (TTP).

[0031] In another embodiment, a computer-implemented method according to the present invention is provided, wherein the prediction is made about a second evaluation time at a first evaluation time, the multi-modal features of the patient are collected at a baseline time point and at the first evaluation time, using the multi-modal features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is made, collected at the baseline time point and at the first evaluation time, and using the result of the treatment response at the second evaluation time, the complementary and predictive machine learning model is trained, and the list of informative feature identifiers is obtained.

[0032] In yet another further embodiment, a computer-implemented method according to the present invention is provided, wherein the prediction of the treatment effect of the patient is represented as a prediction of the response of the patient to the treatment, wherein the patient has cancer, and the treatment is immunotherapy, chemotherapy (e.g., neoadjuvant chemotherapy (NCT), etc.), targeted therapy, treatment using an anti-angiogenic agent, surgery, radiotherapy, or a combination of these treatments.

[0033] In another further embodiment, a computer-implemented method according to the present invention is provided, wherein said prediction of the treatment effect of said patient is represented as a prediction of the response of said patient to said treatment, wherein said patient has lung cancer, and said treatment is immunotherapy, chemotherapy, a combination of immunotherapy and chemotherapy, neoadjuvant therapy, targeted therapy, treatment using an anti-angiogenic agent, surgery, radiotherapy, hyperthermia, and / or adjuvant therapy after surgery, wherein the multimodal features of said patient are the treatment start date for said patient, and the response to treatment at a first evaluation, and the date and progression indicator at a first evaluation, and the date and survival indicator at a first evaluation, and a plurality of clinical features including a plurality of biological features including the expression level of PD-L1 at the baseline time point, a plurality of radiomics features including a plurality of features extracted from said radiological imaging data at the baseline time point and at a first evaluation, a plurality of genomics features including the EGFR gene mutation status and the ALK gene mutation status at the baseline time point, including.

[0034] In another embodiment, a computer-implemented method according to the present invention is provided, wherein the training step of said complementary machine learning model inputs a set of multimodal features including at least two types of features selected from the clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom said prediction is to be made into a machine learning learned by a training algorithm, for each patient in said cohort, at least one of said multimodal features is collected at at least two time points, and said trained complementary machine learning model generates, as an output, a complete list of a plurality of features for a certain patient from an incomplete list.

[0035] In a further embodiment, during the training of the complementary machine learning model, metrics regarding changes between values of at least one multi-modal feature collected at at least two time points for each patient in the cohort are further calculated, at least one longitudinal feature is obtained for each patient in the cohort, and the training step of the complementary machine learning model further includes a step of inputting the at least one longitudinal feature into machine learning learned by a training algorithm.

[0036] In another further embodiment, a computer-implemented method according to the present invention is provided, wherein the prediction of the treatment effect of the patient is represented as a prediction of the response of the patient to the treatment, and the training step of the predictive machine learning includes inputting a set of multi-modal features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and receiving the same treatment, and at least one longitudinal feature, into machine learning learned by a training algorithm, at least one of the multi-modal features is collected for each patient in the cohort at at least two time points, metrics regarding changes between values of at least one multi-modal feature of each patient collected at at least two time points are calculated, at least one longitudinal feature is obtained for each patient, and the trained predictive machine learning model generates as output a label classification of the response of the patient to the treatment or the likelihood of the response of the patient to the treatment and a list of highly informative feature identifiers used for training the predictive machine learning model.

[0037] In another further embodiment, a computer-implemented method according to the present invention is provided, wherein the prediction of the treatment effect of the patient is represented as a prediction of the treatment efficacy of the patient, and the training step of the predictive machine learning inputs into the machine learning a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made, and at least one longitudinal feature, and for each patient in the cohort, at least one of the multimodal features is collected at at least two time points, a metric for the change between the values of at least one multimodal feature of each patient collected at at least two time points is calculated, and at least one longitudinal feature is obtained for each patient. And, the trained predictive machine learning model generates as output a label classification for the treatment efficacy defined as the length of time to an event and a list of informative feature identifiers used for training the predictive machine learning model.

[0038] In another embodiment, a computer-implemented method according to the present invention is provided, wherein the output is supplemented by the list of informative feature identifiers used for training the predictive machine learning model and / or a report with a list regarding the relative contributions of multiple features used in the method for predicting, for example, the treatment response or treatment efficacy of the patient, the treatment effect of the patient.

[0039] In another embodiment, a computer-implemented method according to the present invention is provided, wherein features are complemented based on different feature modalities. In another embodiment, a computer-implemented method according to the present invention is provided, wherein the multimodal features of the patient are at least 75% complete. BRIEF DESCRIPTION OF THE DRAWINGS

[0040]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

[0041] The term "cohort" or "cohort of patients" refers to a group of patients who meet a list of biological and / or clinical criteria. The terms "baseline data", "baseline features", "baseline clinical data, biological data, genomic data and / or radiological data", or "baseline clinical features, biological features, genomic features and / or radiological features" refer to all data or all features collected before the start of treatment. In such context, "baseline" refers to the time before the start of treatment. It may be represented herein as time = 0 or t0.

[0042] The terms "first evaluation (time) data", "first evaluation (time) feature quantity", "first evaluation (time) clinical data, biological data, genomic data and / or radiological data", or "first evaluation (time) clinical feature quantity, biological feature quantity, genomic feature quantity and / or radiological feature quantity" refer to all data or all feature quantities collected after the start of treatment. In such context, "first evaluation" refers to the time after the start of treatment. It may be represented as time = 1 or t1 in this specification. In one embodiment, the first time evaluation is after the start of treatment, preferably about 2 to 4 months after the start of treatment.

[0043] The terms "second / further evaluation (time) data", "second / further evaluation (time) feature quantity", "second / further evaluation (time) clinical data, biological data, genomic data and / or radiological data", or "second / further evaluation (time) clinical feature quantity, biological feature quantity, genomic feature quantity and / or radiological feature quantity" refer to all data or feature quantities collected after the continuation of treatment. In such context, the second / further time evaluation is after the continuation of treatment, preferably about 2 to 4 months after such continuation of treatment (e.g., the second / further administration of the same treatment, change of treatment). It may be represented in this specification as the second evaluation time, i.e., time = 2 or t2, or the third evaluation time, i.e., time = 3 or t3, etc. In one embodiment, the second or further time evaluation is after the continuation of treatment, preferably about 2 to 4 months after the continuation of treatment (e.g., the second / further administration).

[0044] In a preferred embodiment, for one patient, all the said data or feature quantities collected at the baseline time point or at the first evaluation time or at the second evaluation time or at a further evaluation time are assigned to one point of time point, which is represented as the baseline evaluation time or the first evaluation time or the second evaluation time or the further evaluation time respectively in this specification. Thus, even if some of the said data or feature quantities collected for one patient have different collection dates, they are preferably assigned to the same baseline evaluation date or the first evaluation date or the second evaluation date or the further evaluation date.

[0045] In another embodiment, all said data or features collected at the baseline evaluation time or the first evaluation time or the second evaluation time or further evaluation times for a single patient are assigned to their collection dates, which are represented herein as the baseline evaluation time or the first evaluation time or the second evaluation time or further evaluation times, respectively.

[0046] It should be understood that "treatment" or "therapy" refers to any treatment, such as cancer treatment, etc. "First-line treatment" or "primary / initial treatment" or "induction therapy" refers to the first, or primary treatment recommended for a specific disease such as cancer. For example, the first-line treatment for stage IV NSCLC can be single-agent therapy with pembrolizumab, combination therapy of chemotherapy and pembrolizumab, two-drug combination chemotherapy, and any other appropriate treatment. "Second-line treatment" is a treatment for a specific disease such as cancer, which is administered after the failure, interruption, or unacceptable side effects of the said first-line treatment.

[0047] The term "treatment effect" refers to any impact of the treatment on the patient, and may be represented, for example, as the patient's response to the treatment or as the patient's treatment efficacy (defined as the length of time to an event).

[0048] It should be understood that the response of the patient to treatment may be classified into two stages as (1) having a response (complete response or partial response) or (2) having no response (stable disease or progression). Alternatively, the response of the patient to treatment may be classified as (1) complete response, where the symptoms disappear and there are no signs of the disease; (2) partial response, where the symptoms have attenuated by a few percentage points but the disease remains; (3) stable disease, where the symptoms and the disease are not progressing but are not decreasing either; or (4) progression, where the disease is further developing. The response of the patient to cancer treatment may be classified as (1) complete response, where all of the cancer or tumor has disappeared and there are no signs of the disease; (2) partial response, where the cancer has shrunk by a few percentage points but the disease remains; (3) stable disease, where the cancer is neither shrinking nor growing (no change in cancer progression); or (4) progression, i.e., the cancer progresses such that it further develops. Alternatively, the response of the patient to treatment may be provided as the likelihood of response of the patient to the treatment. Additionally, a confidence interval may be provided.

[0049] It should be understood that the therapeutic efficacy of the patient may be defined as the length of time to an event, and examples include progression-free survival (PFS), overall survival (OS), duration of response (DoR), and time to progression (TTP).

[0050] The term "progression-free survival (PFS)" refers to the length of time during and after the cancer treatment period, during which the patient coexists with the disease but the disease is not progressing (evaluated as tumor progression, appearance of new lesions, and / or death).

[0051] The term "overall survival (OS)" starts at the time of diagnosis (or start of treatment) and refers to the length of time until death. PFS and OS are generally referred to as endpoints of survival and are used to measure the efficacy of cancer treatment.

[0052] The term "duration of response (DoR)" refers to the length of time from the response (R) of cancer to treatment (improvement) until the disease worsens again (progression / death). DoR is generally referred to as an early efficacy endpoint.

[0053] The term "time to progression (TTP)" refers to the length of time from the date of diagnosis or the start of treatment for cancer until the cancer worsens or spreads to other parts of the body. It should be understood that PFS, OS, DoR, and TTP are known endpoints for cancer clinical trials, which are time-to-event data, and are used to measure the efficacy of cancer treatment.

[0054] Generally, it should be understood that "data" refers to information directly collected from a patient's health diagnosis, such as medical history, images, blood sample analysis, and genomics tests that provide a list of variations. "Preprocessing" refers to a set of digital processes that lead to the conversion of raw data directly collected from a patient's health diagnosis into a set of "features" that can be used in its machine learning algorithm.

[0055] In addition, "feature preprocessing" or "non-linear feature preprocessing" refers to the use of a set of equations, preferably non-linear equations, to define the metrics / speeds for the changes between a set of features having values collected at one time point (derived from the transformed data of the raw data collected from a patient at one time point) and the same set of features having values collected at another time point (derived from the transformed data of the raw data collected from a patient at another time point).

[0056] It should be understood that a machine learning algorithm takes an input of variable(s) X and outputs variable(s) Y. Completion is to be understood as the process of replacing missing data with substitute values. Therefore, missing values are completed in this completion process. The term "completion" refers to inferring the values of missing data from the available data in an incomplete data set using a statistical analysis model, for example, when the data record is only partially filled and / or some data elements or data features cannot be calculated.

[0057] The term "extraction" refers to signal processing analysis and / or calculation of quantifiable values, for example, features derived from digital data (such as digital data file records, digital health databases, or input of digital signal data), such as, but not limited to: images or one or more image slices, or image elements, or genome sequence files or genomic variant files, or clinically annotated files, or biological laboratory report files. Extraction includes various processing steps, such as, but not limited to: conversion to predetermined units; normalization to scalar values within a range between 0 and 1; mathematical operations such as adding, multiplying, subtracting, dividing, or deriving values from one or more elements in the digital data to measure the nature of the data (such as the development of area or volume); filtering of part or all of the digital data signal, such as cropping, segmenting, extracting subsets such as patterns, regions or volumes of interest, and removing elements such as redundant information; transformation of the digital data signal to different representation spaces (such as from a spatial domain to a transform domain); digital signal processing analysis methods, or combinations thereof.

[0058] The term "aggregation" refers to combining multiple data inputs into a single data display, for example, by concatenation. The term "selection" refers to specifying a subset of multiple data elements into one data set.

[0059] The term "prediction" refers to inferring the value of a result at a future time from the value of a dataset at the current time using a statistical analysis model. A "machine learning model" refers to a data model or data classification that is trained using supervised, semi-supervised, or unsupervised learning techniques, as known in data science, as opposed to an explicit statistical model. This data input can be represented as a 1D signal (vector), 2D signal (matrix), or more generally a multi-dimensional array signal (e.g., a tensor, i.e., an RGB color image represented by three matrices of 3x2D signals for the red, green, and blue color decomposition planes), and / or combinations thereof. A multi-dimensional array is mathematically defined by a data structure arranged along at least two dimensions (each dimension recording values of 1 or more).

[0060] In the case of a deep learning classifier, its data input is further processed through a set of data processing layers, implicitly capturing its hidden data structure, its data features, and underlying patterns. Thanks to the use of multiple data processing layers, deep learning facilitates the generalization of automated data processing for messy and complex pattern detection and data analysis tasks. This machine learning model can be trained within the framework of supervised, semi-supervised, or unsupervised learning. Among supervised learning frameworks, the model learns the function of illustrating output results from an input data set based on exemplary pairs of multiple inputs and corresponding outputs. Examples of machine learning models used for supervised learning include support vector machines (SVMs), regression analysis, linear regression, logistic regression, naive Bayes, linear discriminant analysis, decision trees, k-nearest neighbor algorithms, random forests, artificial neural networks (ANNs) such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), fully connected neural networks, long short-term memory (LSTM) models, and others; and / or combinations thereof. A model trained within an unsupervised learning framework infers a function that identifies the hidden structure of a data set without requiring prior knowledge about the data. Examples of technically known unsupervised machine learning models include clustering (e.g., k-means clustering, mixture model clustering, hierarchical clustering); anomaly detection methods; principal component analysis (PCA), independent component analysis (ICA), t-distributed stochastic neighbor embedding (t-SNE); generative models; and / or unsupervised neural networks; autoencoders; and / or combinations thereof. Semi-supervised learning (SSL) is a machine learning framework that can train a model using both labeled and unlabeled data. Data augmentation methods may optionally be used to generate artificial data samples from a scarce set of real data samples, increasing the quantity and diversity of data used for model training. When combined with a small amount of labeled data, unlabeled data can produce a significant improvement in learning accuracy compared to other frameworks. This approach is particularly interesting when only a portion of the available data is labeled.

[0061] The "Convolutional Neural Network" or "CNN" refers to a machine learning model that uses multiple data processing layers known as convolutional layers to represent its input data in the most suitable way for solving classification or regression tasks. During the training period, in order to perform stochastic gradient descent, optimization algorithms known technically, such as the backpropagation algorithm, are used, and the weight parameters are optimized for each CNN layer. At runtime, the resulting trained CNN may process its input data very efficiently, for example, in the case of a learned classification task, to classify the input data into the correct data output labels with few false positives and false negatives as much as possible. The convolutional neural network may be combined with a recurrent neural network to generate a deep learning classifier.

[0062] Clinical data, biological data, genomic data, and radiological data The term "multimodal data" or "multi-omics data" refers to a set of at least two types of data selected from clinical data, biological data, genomic data, and radiological data. Any combination of at least two types of data selected from clinical data, biological data, genomic data, and radiological data can be represented as multimodal data or multi-omics data. Each data type (clinical data, biological data, genomic data, and radiological data) can be represented as a data modality. Possible sets may thus include clinical data and biological data; clinical data and genomic data; clinical data and radiological data; biological data and genomic data; biological data and radiological data; genomic data and radiological data; clinical data, biological data, and genomic data; clinical data, biological data, and radiological data; biological data, genomic data, and radiological data; clinical data, genomic data, and radiological data; and clinical data, biological data, genomic data, and radiological data.

[0063] In a preferred embodiment, all methods of the present invention are used in combination with at least one data selected from biological data, genomic data, and radiological data, at least clinical data.

[0064] In another preferred embodiment, all methods of the present invention use clinical data, biological data, genomic data, and radiological data. The selection of at least two types (i.e., data modalities) of data (selected from clinical data, biological data, genomic data, and radiological data) for use in the methods of the present invention is based on a cohort of a particular patient group and the prediction target(s), and is similarly based on the evidence currently available regarding their relevance and potential predictive power for predicting response to the selected treatment. Thus, the selection of at least two types of data for use in the methods of the present invention may change over time depending on the newly collected evidence and the predictive power of the response to the selected treatment.

[0065] In a further embodiment, at most three types of data selected from clinical data, biological data, genomic data, and radiological data are used in the methods described herein.

[0066] In one particular embodiment, at least clinical data and biological data are used in the methods described herein, and a computer-implemented method for predicting treatment response or treatment prognosis or treatment efficacy is used for patients suffering from renal cancer.

[0067] In one particular embodiment, at least clinical data, biological data, and radiological data are used in the methods described herein, and a computer-implemented method for predicting treatment response or treatment prognosis or treatment efficacy is used for patients suffering from brain tumors.

[0068] In one particular embodiment, clinical data, biological data, and radiological data are used in the methods described herein, and a computer-implemented method for predicting treatment response or treatment prognosis or treatment efficacy is used for patients suffering from non-small cell lung cancer (NSCLC), particularly patients suffering from stage IV non-small cell lung cancer.

[0069] In one embodiment, multimodal data may be obtained or generated from one or more sources. This may include healthcare providers such as hospitals, primary care units, third-party institutions providing services for medical data analysis and storage, etc. For example, clinical data and biological data may be obtained from databases known by terms such as electronic medical records (EMRs), electronic health records (EHRs), personal health records (PHRs) or electronic case report forms (eCRFs). Clinical data may be obtained from a laboratory information management system (LIMS). Radiological data may be obtained from a picture archiving and communication system (PACS). Genomic data may be obtained from any system storing genomic sequences obtained, for example, through NGS VCF or FastQ files.

[0070] The definition and exemplary content of categories of clinical data, biological data, genomic data and radiological data are described herein. Furthermore, the data is defined, in particular, based on examples of lung cancer.

[0071] Data is selected for a portion of the patients in the cohort, or for the patients for whom the prediction method is executed. It should be understood that when the multimodal data collected for a patient is collected at various time points, they are stored together or separately in a database of one or more patients. Such multimodal patient data can be evaluated at the time of performing the method according to the present invention.

[0072] Clinical data The clinical data of any patient includes, but is not limited to, this patient's: Demographics such as gender, age (date of birth), ethnicity, height, and weight, and medical history such as smoking status, personal medical history, and family medical history, and disease onset date, disease stage, performance status at diagnosis, treatment history, hospitalization and / or death and affected organs, performance status and clinical response at the first / further evaluation time point, progression status (including progression date and site of progression), treatment status after progression, and the latest available vital status, etc., of the medical history, and may be included.

[0073] Clinical data for cancer patients may include, but are not limited to, for this patient: Demographics such as gender, age (date of birth), ethnicity, height, and weight, and medical history such as smoking status, history of autoimmune diseases, underlying diseases, family cancer history, and personal cancer history, and cancer diagnosis date, cancer stage, cancer subtype (IVA, IVB), performance status at diagnosis, history of corticosteroid and antibiotic treatment, treatment methods received and the number of cycles for each progression and / or in total, the presence of treatment toxicity leading to treatment interruption, hospitalization and / or death and affected organs, performance status and clinical response at the first / further evaluation time point, progression status (including progression date and site of progression), treatment status after progression, and the latest available vital status, etc., of the medical history, and may be included.

[0074] Clinical data for patients diagnosed with stage IV NSCLC may include, but are not limited to, for this patient: Demographics such as gender, age (date of birth), ethnicity, height, and weight, and medical history such as smoking status, history of autoimmune diseases, underlying lung diseases, family cancer history, and personal cancer history, and The date of receiving the diagnosis of stage IV NSCLC and the subtype (IVA, IVB), the performance status at the time of the stage IV NSCLC diagnosis, the history of corticosteroid and antibiotic treatment, the treatment methods received and the number of cycles for each progression and / or in total, the presence of treatment toxicity leading to treatment interruption, hospitalization and / or death and the affected organs, the performance status and clinical response at the first / further evaluation points, the progression status (including the date of progression and the site of progression), the treatment status after progression and the latest available vital status, etc., medical history, may be included.

[0075] In one embodiment, the clinical data for a patient suffering from lung cancer, particularly stage IV NSCLC, may include the patient's age, the performance status (PS) of the Eastern Cooperative Oncology Group (ECOG) of the United States, the history of autoimmune diseases, the history of steroid treatment, the state of the gut microbiota, the history of antibiotic treatment, the medical history (e.g., liver metastases, brain metastases and bone metastases), and at least one data regarding immune-related side effects.

[0076] In one embodiment, clinical data for a patient suffering from lung cancer, particularly stage IV NSCLC, may include or consist of at least one of the following pieces of information: gender, age (date of birth), ethnicity, height, weight, smoking status, history of autoimmune diseases, underlying lung diseases, family cancer history, personal cancer history, date of first NSCLC diagnosis, stage of first NSCLC diagnosis, date of stage IV NSCLC diagnosis, subtype (IVA, IVB) of stage IV at diagnosis, performance status at diagnosis, corticosteroid treatment received less than 12 months prior to stage IV NSCLC diagnosis, antibiotic treatment received less than 1 month prior to stage IV NSCLC diagnosis, treatment using pembrolizumab, combination treatment of pembrolizumab and chemotherapy, or treatment using combination chemotherapy of two drugs, start date of treatment, dosing scheme of pembrolizumab, prescription regimen of combination chemotherapy of two drugs, number of cycles of treatment received according to the first evaluation, number of cycles of treatment received with each progression, presence of treatment toxicity leading to treatment interruption, hospitalization or death and affected organs, performance status at the first / further evaluation time point, clinical response at the first / further evaluation time point, latest available progression status, date and site of progression, treatment status after progression, secondary treatment received, latest date and vital status, cause of death.

[0077] Clinical data refers to information that may include descriptive data such as the patient's response to treatment status and the progression of the patient's disease. This descriptive data may be further categorized. For example, the patient's response to treatment may be classified as complete response, partial response, stable disease or progression, where the progression of the patient's disease may be classified as an increase in growth rate and degree of invasion of tumor cells. This categorization may be assigned to the value variables at the preprocessing step.

[0078] Biological data Biological data for a patient may include, but is not limited to, the patient's: disease type and stage, expression level of related receptors, blood analysis (hematological and biochemical analysis) at baseline and at the first / further evaluation time point.

[0079] Biological data for cancer patients may include, but are not limited to, the cancer stage and pathological histological type at the time of diagnosis, the expression level of related receptors, and blood analysis (hematological and biochemical analysis) at the baseline and first / further evaluation time points of the patient.

[0080] Biological data for patients diagnosed with stage IV NSCLC may include, but are not limited to, the pathological histological type of stage IV NSCLC at the time of diagnosis, the expression level of PD-L1, the immunohistochemical antibody used for the measurement of PD-L1, and blood analysis (hematological and biochemical analysis) at the baseline and first / further evaluation time points of the patient.

[0081] In one embodiment, biological data for patients suffering from lung cancer, particularly stage IV NSCLC, may include data on PD-L1 expression on tumor cells. In one embodiment, biological data for patients suffering from lung cancer, particularly stage IV NSCLC, may include at least one data of neutrophil-to-lymphocyte ratio, the level of enzyme lactate dehydrogenase (LDH), and / or blood tumor mutational burden (bTMB).

[0082] In one embodiment, the biological data for a patient suffering from lung cancer, particularly stage IV NSCLC, may include or consist of at least one of the following pieces of information: the histopathological type of stage IV NSCLC at diagnosis, the expression level of PD-L1, the immunohistochemical antibody used for measuring PD-L1, the date of blood analysis at the baseline time point, the white blood cell count at the baseline time point, the neutrophil count at the baseline time point, the lymphocyte count at the baseline time point, the monocyte count at the baseline time point, the eosinophil count at the baseline time point, the basophil count at the baseline time point, the platelet count at the baseline time point, the red blood cell count at the baseline time point, the hemoglobin amount at the baseline time point, the LDH level at the baseline time point, the albumin amount at the baseline time point, the CRP amount at the baseline time point, the date of blood analysis at the first / further evaluation time point, the white blood cell count at the first / further evaluation time point, the neutrophil count at the first / further evaluation time point, the lymphocyte count at the first / further evaluation time point, the monocyte count at the first / further evaluation time point, the eosinophil count at the first / further evaluation time point, the basophil count at the first / further evaluation time point, the platelet count at the first / further evaluation time point, the red blood cell count at the first / further evaluation time point, the hemoglobin amount at the first / further evaluation time point, the LDH level at the first / further evaluation time point, the albumin amount at the first / further evaluation time point, the CRP amount at the first / further evaluation time point.

[0083] The biological data for the patient may include digital pathological data and proteomic data. Genomic data The term "genomic data" refers to the digital representation of genomic information such as DNA sequences. The term "genomic data" may be considered to include "molecular data". In the workflow of next-generation sequencing (NGS) bioinformatics, genomic data may refer to any of the raw nucleotide DNA sequences output from a sequencer (in FASTQ file format), and / or nucleotide sequences assigned by comparison with a reference genome (in BAM or SAM file format), and / or a list of variants output from a variant calling step (in VCF file format), and / or a list of annotated variants output from a variant annotation step.

[0084] A "variant" or "genomic variant" refers to a difference in a genomic sequence relative to a reference sequence. In bioinformatics data processing, a variant is uniquely identified by its position along the direction of a chromosome (chr.pos) and the difference relative to the reference genome (ref, alt) at this position. Variants may include single nucleotide polymorphisms (SNPs) or other single nucleotide variants (SNVs), insertions or deletions (INDELs), copy number variants (CNVs), and may also include large rearrangements, substitutions, duplications, translocations, etc. In the secondary analysis workflow of bioinformatics, a person calling variants may apply variant calling to generate one or more variant calls listed in a variant call file (in VCF format). Germline variants are variants inherited from at least one parent that differ from the wild-type genomic values registered in a reference database and are present in all normal cells of that individual. Somatic variants or somatic mutations or somatic changes are variants caused by genomic changes and are present in one or more somatic cells of that individual, for example, in tumor cells.

[0085] "Mutation" or "mutated gene" refers to a gene for which at least one variant has been identified respectively. In the case of the latter, the "mutated gene state" can be classified as having a mutation or, otherwise, as being normal. This state is routinely used as a biomarker in cancer diagnosis and cancer prognosis. For example, ALK gene mutations or EGFR gene mutations have been shown to have specific relevance to lung cancer.

[0086] "Mutational load" or "mutation load" or "mutation burden" or "mutational burden", or for a tumor, "tumor mutational burden" or "tumor mutational load" or "TMB" refers to a biomarker measured as the number of somatic mutations per megabase of genomic sequence being investigated.

[0087] "MSI state" or "Microsatellite Instability status" or "Micro satellite instability status" refers to the state of genomic changes due to insertions or deletions of a small number of nucleotides within microsatellite repeat regions based on single nucleotide repeats (homopolymers) or a small number of nucleotides (heteropolymers), which is caused by defects in the DNA mismatch repair system. This state is routinely used as a biomarker in cancer diagnosis and cancer prognosis, and is particularly used in uterine cancer, colon cancer, and gastric cancer such as UCES (Uterine Corpus Endometrial Carcinoma), COAD (colorectal adenocarcinoma), and STAD (gastric fundic gland carcinoma). The MSI state of genomic changes for a patient is usually categorized as follows: - Microsatellite stable, MSS: A state without any basis among the genomic loci of biomarkers showing instability. - Microsatellite instability - low, MSI-L: A state where there is evidence of instability at only one marker locus. - Microsatellite instability - high, MSI-H: A state where there is evidence of instability at at least two marker loci.

[0088] "Deficiency in homologous recombination state", that is, "HRD state", refers to the classification of the homologous recombination pathway, and is related to any cellular state / event that results in a deficiency in the homologous recombination pathway. The HRD state may be classified as positive (HRD+) if there is a defect in the homologous recombination pathway, negative (HRD-) if there is no defect in the homologous recombination pathway, or may be classified as other with undetermined (HRD undetermined, HRD unknown).

[0089] "Genomic pathway" or "gene pathway" refers to a set of genomic loci or genomic expression that is significantly affected in a specific state. In bioinformatics, the secondary analysis workflow pathway analysis approach uses the available pathway database and specific genomic data or gene expression data from this patient to identify the presence or absence of a genomic pathway for the patient.

[0090] The genomic data for the patient may include, but is not limited to, the mutation status of the patient's disease site, such as obtained through an NGS VCF file. The patient's disease site may be a tumor substance or genetic material released by the tumor and observed in the blood.

[0091] The genomic data of the cancer patient may include, but is not limited to, the mutation status of the patient's cancer, such as obtained through an NGS VCF file. In one embodiment, genomic data for a patient with lung cancer, particularly stage IV NSCLC, may include at least one of the following or may be composed of these: EGFR gene mutation status and ALK gene mutation status. The mutation status of the tumor can be obtained by any known method, such as through an NGS VCF file (based on a locally available NGS panel), Sanger sequencing, immunohistochemistry, etc.

[0092] In one embodiment, genomic data for a patient with lung cancer, particularly stage IV NSCLC, may include data regarding tumor cell mutations, such as mutation status of at least one of the EGFR gene, ALK gene, KRAS gene, STK11 / LKB1 gene, KEAP1 gene, PTEN gene, PIK3CA gene, TP53 gene, ROS1 gene, BRAF gene, NTRK1 / 2 / 3 genes, which are constituent genes of DNA repair pathways such as mismatch repair genes, and loss of functional mutations in POLE gene and BRCA2 gene, which are constituent genes of interferon gamma (IFN-gamma) signaling including JAK1, JAK2, and beta-2-microglobulin (B2M).

[0093] In one embodiment, genomic data for a patient with lung cancer, particularly stage IV NSCLC, may include data regarding tumor immunogenicity, such as tumor mutation burden (TMB), microsatellite instability (MSI), and mismatch repair deficiency (dMMR).

[0094] Genomic data may be collected at the baseline time point or only once at the first evaluation or further evaluations, or may be collected at multiple time points. Radiological / radiomics data Radiological data, also referred to as radiomics data, are the collected images, including but not limited to computed tomography (CT), positron emission tomography (PET), PET / CT, magnetic resonance imaging (MRI), single photon emission computed tomography (SPECT), etc.

[0095] Radiological data for a patient may include, but is not limited to, imaging of the patient at a pre-baseline time point, at a baseline time point, and at a first / further evaluation time point. The terms "medical image data" or "radiological data" or "imaging data" refer to digital image data and one or more images collected for a patient at any time point during the diagnostic and treatment periods. These images can be obtained from examinations of patients at one or more medical centers, which are responsible for one or more imaging modalities such as CT, PET, MRI, SPECT, and others. These images may be in 2D or 3D format. These images can be safely collected, stored, archived, and transmitted to the radiomics processing system of the present invention in accordance with the PACS (Picture Archiving and Communication System) and DICOM (Digital Imaging and Communications in Medicine) digital medical imaging technical standards widely used in healthcare institutions around the world.

[0096] Radiological data for a cancer patient may include, but is not limited to, for the patient: imaging at a pre-baseline time point (if available) (CT scan of the post-injection cancer site in millimeters in the portal phase, section <3 mm), imaging at a baseline time point (CT scan of the post-injection cancer site in millimeters in the portal phase, section <3 mm; if available, PET / CT, CT, MRI), imaging at a first / further evaluation time point (CT scan of the post-injection cancer site in millimeters in the portal phase, section <3 mm; if available, CT, MRI), imaging during the follow-up visit period after the first / further evaluation if available, imaging at progression (CT scan of the post-injection cancer site in millimeters in the portal phase, section <3 mm; if available, PET / CT, CT, MRI), the number of metastatic tumors at each metastatic site at the baseline time point and at the first / further evaluation time point, and if available, the RECIST (Response Evaluation Criteria in Solid Tumors) criteria.

[0097] In one embodiment, the radiological data for a patient suffering from lung cancer, particularly stage IV NSCLC, may include data regarding an imaging-based assessment of the clinical tumor burden.

[0098] The radiological data for the patient diagnosed with stage IV NSCLC may include, but is not limited to, for the patient: imaging at the pre-baseline time point if available (portal phase, millimeter-scale, post-injection chest, abdomen, and lumbar CT scans, slices <3 mm), imaging at the baseline time point (portal phase, millimeter-scale, post-injection chest, abdomen, and lumbar CT scans, slices <3 mm; if available, PET / CT, brain CT, brain MRI), imaging at the first / further evaluation time point (portal phase, millimeter-scale, post-injection chest, abdomen, and lumbar CT scans, slices <3 mm; if available, brain CT, brain MRI), imaging during follow-up visits after the first / further evaluation time point if available, imaging at the progression time point (portal phase, millimeter-scale, post-injection chest, abdomen, and lumbar CT scans, slices <3 mm; if available, brain CT, brain MRI), the number of metastatic tumors at each metastatic site at the baseline time point and the first / further evaluation time point, and RECIST criteria if available.

[0099] In one embodiment, the radiological data for a patient with lung cancer, particularly stage IV NSCLC, may include, or may be composed of, at least one of the following: availability and date of a baseline pre-timepoint chest CT scan, computerized tomography of the thorax at the baseline timepoint, date of abdominal and lumbar (CT / TAP) scans, CT-TAP RECIST at the baseline timepoint (if available), availability and date of a brain CT scan at the baseline timepoint (if available), availability and date of a PET / CT scan at the baseline timepoint (if available), availability and date of a brain MRI at the baseline timepoint (if available), assessment of the degree of metastatic burden by imaging at the baseline timepoint, degree of metastatic disease - status of affected organs at the baseline timepoint, date of chest CT scan at the first / further assessment timepoint, chest CT scan RECIST at the first / further assessment timepoint (if available), availability and date of a CT-TAP scan at the first / further assessment timepoint (if available), availability and date of a brain CT scan at the first / further assessment timepoint (if available), availability and date of a brain MRI at the first / further assessment timepoint (if available), date of follow-up chest CT scan after the first / further assessment timepoint (if available), chest CT scan RECIST of the follow-up chest CT scan after the first / further assessment timepoint (if available), availability and date of a follow-up CT-TAP scan after the first / further assessment timepoint (if available), availability and date of a follow-up brain CT scan after the first / further assessment timepoint (if available), date of chest CT scan at the progressive assessment timepoint, CT scan RECIST at the progressive assessment timepoint (if available), availability and date of a CT-TAP scan at the progressive assessment timepoint (if available), and availability and date of a brain CT scan at the progressive assessment timepoint (if available).

[0100] Pretreatment and clinical, biological, genomic, and radiological features In one embodiment, the acquired data is preprocessed in steps including any known techniques related to data extraction and data preparation so as to be further input to and processed by a machine learning algorithm. It should be understood that a feature refers to a feature value.

[0101] The preprocessing steps may include steps such as recording data variables such as category attributes, calculating the gap time between certain events by a computer, and performing descriptive statistics on the data (minimum value, maximum value, average, median, etc. for value variables, counts, and frequencies of category variables).

[0102] The preprocessing steps may include extracting data from a patient's data platform, such as from the genomic module of SOPHiA DDM (trademark) as a platform and / or from the radiology module of SOPHiA DDM (trademark) as a platform.

[0103] In this step, extraction and harmonization of radiological images can be performed. The preprocessing steps for radiological images may include histogram matching, GAN (adversarial generation) network, filtering, etc.

[0104] In this step, secondary genomic analysis can be performed. The term "multimodal feature" or "multi-omics feature" refers to a set of at least two types of features selected from clinical features, biological features, genomic features, and radiological features. Any combination of at least two types of features of clinical features, biological features, genomic features, and radiological features can be presented as a multimodal or multi-omics feature. Each feature type (clinical feature, biological feature, genomic feature, and radiological feature) can be represented as a feature modality. Thus, possible sets include or consist of: clinical features and biological features; clinical features and genomic features; clinical features and radiological features; biological features and genomic features; biological features and radiological features; genomic features and radiological features; clinical features, biological features, and genomic features; clinical features, biological features, and radiological features; biological features, genomic features, and radiological features; clinical features, genomic features, and radiological features; and clinical features, biological features, genomic features, and radiological features.

[0105] In a preferred embodiment, all methods of the present invention are used in combination with at least one feature type selected from clinical features, biological features, genomic features, and radiological features, at least the clinical features.

[0106] In another preferred embodiment, all methods of the present invention use clinical features, biological features, genomic features, and radiological features. The selection of at least two types of features (i.e., feature modalities) for use in the methods of the present invention (selected from clinical features, biological features, genomic features, and radiological features) is based on a particular prediction target (one or more), and is similarly based on currently available evidence regarding their relevance and potential predictive power for predicting the response to the selected treatment.

[0107] In one particular embodiment, at least clinical features and biological features are used in the manner described herein, and a computer-implemented method for predicting treatment response or treatment prognosis or treatment efficacy is used for patients suffering from kidney cancer.

[0108] In one particular embodiment, at least clinical features, biological features and radiological features are used in the manner described herein, and a computer-implemented method for predicting treatment response or treatment prognosis or treatment efficacy is used for patients suffering from brain tumors.

[0109] In one particular embodiment, clinical features, biological features, genomic features and radiological features are used in the manner described herein, and a computer-implemented method for predicting treatment response or treatment prognosis or treatment efficacy is used for patients suffering from non-small cell lung cancer (NSCLC), particularly stage IV NSCLC.

[0110] In one embodiment, the multimodal features can be obtained, acquired, or generated from one or more sources. In another embodiment, the multimodal features can be separately obtained, acquired, or generated from separate sources.

[0111] A set of clinical features, biological features, genomic features and / or radiological features has a form compatible with a machine learning model. It should be understood that the multimodal features for a patient are obtained from the multimodal data of the patient collected at various time points and are stored together or separately in a database of one or more patients. These multimodal patient features can be evaluated at the time of performing the method according to the present invention.

[0112] Clinical features In one embodiment, the clinical data described herein is preprocessed to obtain clinical features.

[0113] In one embodiment, clinical data refers to information that may include descriptive data such as a patient's response to a treatment situation, the progression of the patient's disease, etc. This descriptive data may be further categorized / classified and assigned to value variables in a preprocessing step.

[0114] In one embodiment, clinical data refers to information that may be a date (e.g., the start date of treatment). This data may be converted into feature quantities in the preprocessing step according to known methods. For example, the date may be converted into a vector of several feature values.

[0115] Biological feature quantities In one embodiment, biological data as described herein is preprocessed to obtain biological feature quantities.

[0116] In one embodiment, biological data refers to information that may include descriptive data such as a patient's disease type and stage, the expression level of related receptors, blood analysis, etc. This descriptive data may be further categorized / classified and assigned to variables in the preprocessing step.

[0117] This data may be converted into certain feature quantities in the preprocessing step according to any known method, such as the step of normalizing (matching with a reference) blood parameters or constructing an index such as the neutrophil-to-lymphocyte ratio, and then converting it into a vector of several feature values.

[0118] Genomic feature quantities In one embodiment, genomic data as described herein is preprocessed to obtain genomic feature quantities.

[0119] In one embodiment, genomic data may include the state of the patient's disease and the state of mutations as described herein. This state of mutations may be preprocessed by any known method to obtain a set of genomic feature quantities.

[0120] In one embodiment, the pre - processing of genomic data may include secondary analysis steps of the genome or tertiary analysis steps of the genome according to known methods using known systems as described herein.

[0121] Radiological features In one embodiment, radiomic metrics or radiological features are extracted from radiological images.

[0122] In one embodiment, radiological data is an image as described herein, and the aforementioned image is pre - processed to obtain a set of adjusted radiological features according to known methods and systems.

[0123] The term "radiomics", which is an abbreviation of "radiology omics", refers to the high - throughput digital extraction of mineable quantitative data from radiological image data.

[0124] The terms "radiomics feature", "image metric", "image descriptor", or "image feature" refer to an imaging biomarker that can be extracted from imaging data as a quantitative summary of the image, such as a statistical value. A radiomics feature or radiomics descriptor refers to a set of values calculated by a computer using the intensity values of 2D pixels or 3D voxels contained in a fragmented ROI or VOI from a fragment of an image region (2D region of interest (ROI) or 3D volume of interest (VOI)). In a radiomics workflow, multiple features, each representative of a different property of the ROI or VOI within the image, can be individually extracted using computer-implemented methods and combined to produce a set of features or radiomics signature. This set of features can include, but is not limited to, morphological features, heterogeneity features, and texture features. Examples of multiple sets of features commonly used in radiomics include any of the well-established features of the 169 IBSI (Image Biomarker Standardization Initiative) criteria, any of the 1500 features of the open-source software package Pyradiomics, and / or any of the public software tools LIFEx (www.lifexsoft.org), CERR, or IBEX.

[0125] Method for predicting a treatment effect (treatment response or treatment efficacy) In one embodiment, a computer-implemented method for predicting a treatment effect of a patient according to the invention described herein is provided. The treatment effect may be expressed as a prediction of the patient's response to the treatment or as the treatment efficacy of that patient. Thus, any embodiment of a computer-implemented method for predicting a treatment response or treatment efficacy should be considered an example of a computer-implemented method for predicting a treatment effect of a patient.

[0126] In one embodiment, a computer-implemented method for predicting a treatment effect of a patient is provided based on a multimodal feature of the patient comprising at least two types of features selected from clinical features, biological features, genomic features, and radiological features.

[0127] In one embodiment, a computer-implemented method is provided for predicting the treatment effect of a patient on a multimodal feature of the patient, including at least two types of features selected from clinical features, biological features, genomic features, and radiological features (predicting the treatment response or treatment efficacy of the patient based on such features).

[0128] In one embodiment, a computer-implemented method is provided for predicting the treatment response or treatment efficacy of a patient based on a multimodal feature of the patient, including at most three types of features selected from clinical features, biological features, genomic features, and radiological features.

[0129] In the method of the present invention, at least one multimodal feature for a patient can be collected at at least two or at least three time points. In a further embodiment, all of the selected multimodal features for a patient are collected at at least two or at least three time points.

[0130] In a further embodiment, a computer-implemented method is provided for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient) based on multimodal data of the patient, comprising obtaining at least two types of data selected from clinical data, biological data, genomic data, and radiological data. This computer-implemented method for predicting the treatment effect of the patient (predicting the treatment response or treatment efficacy of the patient) starts from the patient's data and thus includes the following additional steps: obtaining multimodal data of the patient, including at least two types of data selected from clinical data, biological data, genomic data, and radiological data; Steps for preprocessing the acquired data and steps for obtaining the multi-modal features of the patient. The acquired features can be further processed as described herein for a computer-implemented method of predicting the treatment effect of the patient (predicting the treatment response or treatment efficacy of the patient) based on the multi-modal features of the patient.

[0131] It should be understood that the selection of patient data can be the same as the selection of patient features as described herein. In a preferred embodiment, there is provided a computer-implemented method of predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient), the method comprising the step of obtaining at least the clinical features of the patient combined with at least one type of patient features selected from biological features, genomic features, and radiological features.

[0132] In another preferred embodiment, there is provided a computer-implemented method of predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient), the method comprising the step of obtaining at least the clinical features and biological features of the patient.

[0133] In another preferred embodiment, there is provided a computer-implemented method of predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient), the method comprising the step of obtaining at least the clinical features, biological features, and radiological features of the patient.

[0134] In another preferred embodiment, there is provided a computer-implemented method of predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient), the method comprising the step of obtaining the clinical features, biological features, genomic features, and radiological features of the patient.

[0135] Regardless of whether the patient's multimodal features are defined as including / consisting of any combination of at least two types of features selected from clinical features, biological features, genomic features, and radiological features, it should be understood that the steps of the computer-implemented method of the present invention are performed in the same manner as described herein.

[0136] In a preferred embodiment, there is provided a computer-implemented method for predicting a patient's treatment effect (predicting a patient's treatment response or treatment efficacy) based on a multimodal feature of a patient including at least two features selected from clinical features, biological features, genomic features, and radiological features, wherein at least one multimodal feature for the patient is collected at at least two or at least three time points. In another preferred embodiment, there is provided a computer-implemented method for predicting a patient's treatment effect (predicting a patient's treatment response or treatment efficacy) based on a multimodal feature of a patient including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein at least one multimodal feature for the patient is collected at at least two or at least three time points, and a metric regarding a change between values of at least one multimodal feature for the patient collected at at least two or at least three time points is calculated to obtain at least one longitudinal feature. This calculation of at least one longitudinal feature can be performed before or after the step of inputting the missing patient features (Embodiment 1 or 4). This step regarding the calculation of at least one longitudinal feature performed before the step of inputting the missing patient features can be described as being performed after receiving the multimodal features of the patient and / or before the step of aggregating multiple features into an incomplete feature value vector, as shown in FIG. 4. This step regarding the calculation of at least one longitudinal feature performed after the step of complementing the missing patient features can be described as being performed after obtaining a complete feature value vector and / or before the step of aggregating features into a completely aggregated multimodal longitudinal feature value vector, as shown in FIG. 5.

[0137] The method of the present invention may use the calculation of a "metric for change", i.e., "rate of change (ROC)", between the values of at least one multimodal feature of a patient collected at at least two or at least three time points. As a result of this calculation, at least one longitudinal feature is obtained, which may also be represented herein as a multimodal longitudinal feature. The longitudinal feature has a value related to the change of the feature within a period, for example, A L has an identifier such as

[0138] In one embodiment, a metric for the change between the values of at least one multimodal feature of a patient collected at at least two time points is calculated according to any known linear form.

[0139] In one embodiment, a metric for the change between the values of at least one multimodal feature of a patient collected at at least two time points is calculated according to Equation 1:

[0140]

Equation

[0141] (Equation 1) where A T1 — the value of feature A at T1, A T2 — the value of feature A at T2, T1 — evaluation time 1, T2 — evaluation time 2, A L — the rate of change of the feature value A, that is, the longitudinal feature A. In one embodiment, a metric for the change between at least one multimodal value of a patient collected at at least three time points is calculated according to any known interpolation formula.

[0142] For example, a metric for change can be calculated for a biological feature such as a blood analysis, where, for example, the value of the white blood cell count (B1 (白血球数)t0 ) of a patient at the baseline time point and the white blood cell count (B1 (白血球数)t1A metric regarding the change with respect to the value of (白血球数)L is calculated according to Equation 1, and a longitudinal feature quantity B1 (腫瘍の体積)t0 is obtained. For example, the metric regarding the change can be calculated for a radiological feature quantity such as the volume of a tumor, where, for example, the value of the volume of the patient's tumor at the baseline time point (R1 (腫瘍の体積)t1) ) and the value of the volume of the patient's tumor at the first evaluation time point (R1 (腫瘍の体積)L ), and a metric regarding the change therebetween is calculated according to Equation 1, and a longitudinal feature quantity R1

[0143] is obtained.

[0144] Therefore, a computer-implemented method for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of a patient) can be based on a multimodal feature quantity of the patient including at least two types of feature quantities selected from clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities, and is similar to at least one longitudinal feature quantity calculated as described herein. Both the multimodal feature quantity and the longitudinal feature quantity of the patient can be represented as the feature quantity of the patient.

[0144] In one embodiment, a computer-implemented method for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of a patient) is provided, and the method includes the following: Receiving a multimodal feature quantity of the patient including at least two types of feature quantities selected from clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities, where, optionally, at least one multimodal feature quantity of the patient is collected at at least two time points; Optionally, calculating a metric regarding the change between the values of at least one multimodal feature quantity of the patient collected at at least two time points, and obtaining at least one longitudinal feature quantity; Inputting the multimodal features of the patient into a first trained machine learning model, optionally together with at least one longitudinal feature; and inputting the output of this first trained machine learning model into a second trained machine learning model, optionally together with at least one longitudinal feature, where the second machine learning model uses a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is being made, and at least one of the multimodal features has been collected for each patient in the cohort at at least two time points, and optionally at least one longitudinal feature has been obtained, to predict the treatment effect (treatment response or treatment efficacy) of the patient. Predicting the treatment effect (treatment response or treatment efficacy) of the patient.

[0145] In one embodiment, a computer-implemented method is provided for predicting the treatment effect (treatment response or treatment efficacy) of a patient using at least two trained machine learning models. These models can be referred to as the first trained machine learning model or the second trained machine learning model depending on the order of use in the method of the present invention.

[0146] In one embodiment, a machine learning model is trained to complement missing features of a patient and can be referred to as a complementary machine learning model. In another embodiment, another machine learning model is trained to predict the treatment effect (treatment response or treatment efficacy) of a patient and can be referred to as a predictive machine learning model.

[0147] In another embodiment, the method of the present invention uses a list of feature identifiers identified during the training of a predictive machine learning model, which may be referred to as highly informative feature identifiers. Examples of identifiers include indices, labels, and the like. It should be understood that the step of selecting highly informative feature identifiers is a common process when developing a predictive model and may be performed by any known method. Once longitudinal features are calculated, it should be understood that they may be selected as highly informative features and included in the list of feature identifiers. Both the patient's multimodal highly informative feature identifiers and the highly informative longitudinal feature identifiers can be represented as feature identifiers.

[0148] Figure 1 shows a list of highly informative feature identifiers, where highly clinically informative feature identifiers are exemplified as C1 - C3 at t0, C1 - C4 at t1, and C1 - C3 and C5 at t2; highly informative biological feature identifiers are exemplified as B1 - B3 at t0, B1 - B4 at t1, and B1 - B3 and B5 at t2; highly informative radiomics identifiers are exemplified as R1 - 7 at t0, t1, and t2; and highly informative genomic feature identifiers are exemplified as G1 - G3 at t0, G1 at t1, and G1 and G4 at t2. It should be understood that each highly informative feature identifier identifies a feature having a value. The selected set of highly informative feature identifiers may be the same or different for various time points. For example, as shown in Figure 1, the highly informative clinical feature identifiers C1 - C3 are present at t0, t1, and t2, while C4 is present only at t1 and C5 is present only at t2.

[0149] In one embodiment, the trained machine learning model may be a logistic regression, random forest, support vector machine (also referred to as SVM, support vector network), gradient boosting method, mysvm forest, or any equivalent model. In one embodiment, the trained machine learning model is a supervised machine learning algorithm.

[0150] In one embodiment, a machine learning model is provided for use in complementing a patient's missing feature quantity for use in a computer-implemented method of predicting a patient's treatment effect (predicting a patient's treatment response or treatment efficacy) according to the present invention, wherein a training step of the machine learning model is performed on a machine learning teacher algorithm with a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made. In another embodiment, for training a machine learning model that complements a patient's missing feature quantity, at least one multimodal feature quantity for each patient in the cohort may be collected at at least two time points or at least three time points. In another embodiment, for training a machine learning model that complements a patient's missing feature quantity, at least one multimodal feature quantity for each patient in the cohort may be collected at at least two time points or at least three time points, and in addition, a metric for the change between the values of at least one multimodal feature quantity of each patient collected at at least two time points or at least three time points may be calculated, and as a result, at least one longitudinal feature quantity may be obtained for each patient. In a further embodiment regarding the training of a machine learning model that complements a patient's missing feature quantity, all selected multimodal feature quantities collected for the patients in the cohort are collected at at least two time points or at least three time points, and optionally, a metric for the change between the values of the multimodal feature quantities of each patient collected at at least two time points or at least three time points may be calculated, and as a result, longitudinal feature quantities may be obtained for each patient.In a further embodiment regarding the training of a machine learning model for complementing the missing feature amounts of a patient, all the selected multimodal feature amounts collected for the patients of the cohort are collected at at least two time points or at at least three time points, and metrics regarding the changes between the values of the multimodal feature amounts of each patient collected at at least two time points or at at least three time points are calculated, and as a result longitudinal feature amounts are obtained for each patient.

[0151] In one embodiment of the method of the present invention, when it is executed before the step of complementing the multimodal feature amounts of a patient for which the calculation of at least one longitudinal feature amount is missing and the at least one longitudinal feature amount, the multimodal feature amounts and the at least one longitudinal feature amount of each patient of the cohort are used for the training of the complementing machine learning model (for example, FIG. 4, Embodiment 2 or 5).

[0152] In an alternative embodiment of the method of the present invention, when it is executed before the step of complementing the multimodal feature amounts of a patient for which the calculation of at least one longitudinal feature amount is missing, the multimodal feature amounts of each patient of the cohort are used for the training of the complementing machine learning model.

[0153] In the method of the present invention, when it is executed after the step of complementing the multimodal feature amounts of a patient for which the calculation of at least one longitudinal feature amount is missing, the multimodal feature amounts of each patient of the cohort are used for the training of the complementing machine learning model (for example, FIG. 5, Embodiment 3 or 6).

[0154] A machine learning model for complementing the missing feature amounts of a patient is trained to output a complete list of a plurality of feature amounts for a certain patient from an incomplete list. In one embodiment of the method of the present invention, when executed before the step of complementing the multi-modal features of a patient lacking the calculation of at least one longitudinal feature with at least one longitudinal feature, the machine learning model for complementing the missing features of the patient outputs a complete list of features including the multi-modal features and at least one longitudinal feature (i.e., the features of the patient) for the patient from an incomplete list of features (e.g., FIG. 4, Embodiment 2 or 5). As an optional further step, the missing multi-modal features of the patient from one time point may be inferred based on the complete list of longitudinal features obtained.

[0155] In an alternative embodiment of the method of the present invention, when executed before the step of complementing the multi-modal features of a patient lacking the calculation of at least one longitudinal feature, the machine learning model for complementing the missing features of the patient is trained to output a complete list of features including the multi-modal features for a patient from an incomplete list including the multi-modal features. As an optional further step, at least one longitudinal feature may be calculated as described herein based on the complete list of features including the multi-modal features obtained.

[0156] In the method of the present invention, when executed after the step of complementing the multi-modal features of a patient lacking the calculation of at least one longitudinal feature, the machine learning model for complementing the missing features of the patient is trained to output a complete list of features including the multi-modal features for the patient from an incomplete list including the multi-modal features (e.g., FIG. 5, Embodiment 3 or 6).

[0157] In another embodiment, there is provided a machine learning model for complementing missing features of a patient for use in a computer-implemented method for predicting a treatment effect (predicting a treatment response or treatment efficacy) of a patient according to the present invention, wherein the training step of the complementing machine learning model is a machine learning teacher-aided training algorithm, and inputs a set of multi-modal features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made, and for each patient in the cohort, at least one of the multi-modal features is collected at at least two time points, and the trained complementing machine learning model generates, as an output, a complete list of features for a patient from an incomplete list.

[0158] In yet another further embodiment, there is provided a machine learning model for complementing missing features of a patient for use in a computer-implemented method for predicting a treatment effect (predicting a treatment response or treatment efficacy) of a patient according to the present invention, wherein the training step of the complementing machine learning model is a machine learning teacher-aided training algorithm and includes the step of inputting at least one longitudinal feature obtained for each patient in the cohort as described herein.

[0159] In another embodiment, there is provided a machine learning model for predicting a treatment effect of a patient for use in a computer-implemented method for predicting a treatment effect of a patient according to the present invention, wherein the training step of the machine learning model is a machine learning teacher-aided training algorithm and includes the step of inputting a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made. Exemplary embodiments for the training of the prediction machine learning model for use in a method for predicting a treatment effect (predicting a treatment response or treatment efficacy) of a patient are described below in this specification.

[0160] In another embodiment, a machine learning model for predicting a patient's treatment response is provided for use in a computer-implemented method of predicting a patient's treatment response according to the present invention, wherein the training step of the machine learning model comprises inputting a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made, using a supervised learning algorithm.

[0161] In another embodiment, for training a machine learning model for predicting a patient's treatment response, at least one multimodal feature may be collected for each patient in the cohort at at least two time points or at least three time points. In another embodiment, at least one of these multimodal features may be collected for each patient in the cohort at at least two time points or at least three time points, where metrics for changes between values of at least one multimodal feature of each patient collected at at least two time points or at least three time points may be calculated, such that for each patient at least one longitudinal feature is obtained, and this at least one longitudinal feature is used in addition to the training of the machine learning model for predicting the patient's treatment response. In a further embodiment for training a machine learning model for predicting a patient's treatment response, all selected multimodal features collected for the patients in the cohort are collected at at least two time points or at least three time points, and optionally, metrics for changes between values of the multimodal features of each patient collected at at least two time points or at least three time points are calculated, such that longitudinal features may be obtained for each patient. These embodiments may be used in the method of the present invention when performed before or after the step of complementing the multimodal features of patients lacking the calculation of at least one longitudinal feature (Embodiment 1).

[0162] In another embodiment, there is provided a machine learning model for predicting a patient's treatment response for use in a computer-implemented method of predicting a patient's treatment response according to the present invention, wherein the training step of this predictive machine learning includes clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patients for whom the prediction is being made, selected from at least two types of features. The method includes inputting a set of multimodal features including at least two types of features and at least one longitudinal feature, collecting at least one of the multimodal features for each patient in the cohort at at least two time points, calculating a metric for the change between the values of at least one multimodal feature collected for each patient at at least two time points, then obtaining at least one longitudinal feature for each patient, and the trained predictive machine learning model generates as output a labeled classification of the patient's response to the treatment or the likelihood of the patient's response to the treatment, and a list of highly informative feature identifiers used for training the predictive machine learning model.

[0163] The machine learning model for predicting a patient's treatment response is trained to generate as output a labeled classification of the patient's response to the treatment or the likelihood of the patient's response to the treatment, and a list of highly informative feature identifiers used for training the predictive machine learning model. As used herein, this list of highly informative feature identifiers may include highly informative clinical feature identifiers, highly informative biological feature identifiers, highly informative genomic feature identifiers, and / or highly informative radiological feature identifiers and / or highly informative longitudinal feature identifiers.

[0164] In another embodiment, there is provided a machine learning model for predicting a patient's treatment efficacy for use in a computer-implemented method of predicting a patient's treatment efficacy according to the present invention, wherein the training step of this machine learning model is based on a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made, and includes inputting a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features. In another embodiment, for training the machine learning model for predicting a patient's treatment efficacy, for each patient in the cohort, at least one multimodal feature can be collected at at least two time points or at least three time points. In another embodiment, for each patient in the cohort, at least one of the multimodal features may be collected at at least two time points or at least three time points, and metrics for changes between the values of at least one multimodal feature of each patient collected at at least two time points or at least three time points may be calculated, whereby for each patient at least one longitudinal feature is obtained, and this at least one longitudinal feature is additionally used in training the machine learning model for predicting a patient's treatment efficacy. In a further embodiment for training the machine learning model for predicting a patient's treatment efficacy, all selected multimodal features collected for the patients in the cohort are collected at at least two time points or at least three time points, and optionally, metrics for changes between the values of the multimodal features of each patient collected at at least two time points or at least three time points are calculated, whereby longitudinal features may be obtained for each patient. These embodiments can be used in the method of the present invention when executed before or after the step of complementing the multimodal features of patients lacking the calculation of at least one longitudinal feature (Embodiment 4).

[0165] In another embodiment, there is provided a machine learning model for predicting a patient's treatment efficacy for use in a computer-implemented method of predicting a patient's treatment efficacy according to the present invention. The training step of this predictive machine learning includes inputting a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patients for whom the prediction is to be performed by a machine learning teacher algorithm and at least one longitudinal feature. For each patient in the cohort, at least one of these multimodal features is collected at at least two time points, then a metric for the change between the values of at least one multimodal feature of each patient collected at at least two time points is calculated, and at least one longitudinal feature of each patient is obtained. And this trained predictive machine learning model generates as output a label classification of treatment efficacy defined as the length of time to an event and a list of information identifiers used for the training of the predictive machine learning model.

[0166] The machine learning model for predicting a patient's treatment efficacy is trained to generate as output a label classification of treatment efficacy defined as the length of time to an event and a list of highly informative feature identifiers used for the training of the predictive machine learning model. As used herein, the list of highly informative feature identifiers may include highly informative clinical feature identifiers, highly informative biological feature identifiers, highly informative genomic feature identifiers, and / or highly informative radiological feature identifiers and / or highly informative longitudinal feature identifiers.

[0167] In one embodiment, a computer-implemented method is provided for predicting the treatment effect (predicting the treatment response or treatment efficacy) of a patient based on the patient's multimodal features, where the features are obtained or received from an external database such as a database storing preprocessed patient data. In one embodiment, the patient's multimodal features are obtained or received separately. In another embodiment, the longitudinal features of the patient are obtained or received from an external database such as a database storing preprocessed patient data, and metrics for changes as described herein may be calculated.

[0168] The multimodal features of the patient for use in the method of the present invention (e.g., received in step b) should be selected from the multimodal features of a cohort of patients used to train a complementary and predictive machine learning model (e.g., obtained in step a), although the multimodal features of the patient are not fully available (thus, the set of features is not complete). This cohort of patients (whose data is used for training the machine learning model) has the same disease and has received the same treatment as the patient for whom the prediction is being made.

[0169] In one embodiment, a computer-implemented method is provided for predicting the treatment effect (predicting the treatment response or treatment efficacy) of a patient based on the patient's multimodal features, where the features are not complete; in other words, the obtained or received features are partial.

[0170] In one embodiment, the multimodal features of the patient provided as an input to a computer-implemented method for predicting the treatment effect (predicting the treatment response or treatment efficacy) of a patient are not complete, and the method includes a step of complementing missing features.

[0171] In one embodiment, in the method of the present invention performed before the step of complementing the multimodal features of a patient with at least one longitudinal feature missing, clinical missing features, biological missing features, genomic missing features and / or radiological missing features and longitudinal missing features are complemented (e.g., FIG. 4, Embodiment 2 or 5). As an optional further step, the missing multimodal features of the patient from one time point may be inferred based on the obtained longitudinal features.

[0172] In an alternative embodiment, in the method of the present invention performed before the step of complementing the multimodal features of a patient with at least one longitudinal feature missing, clinical missing features, biological missing features, genomic missing features and / or radiological missing features are complemented, and optionally longitudinal missing features are calculated therefrom (e.g., FIG. 4, Embodiment 2 or 5).

[0173] In another embodiment, in the method of the present invention performed after the step of complementing the multimodal features of a patient with at least one longitudinal feature missing, clinical missing features, biological missing features, genomic missing features and / or radiological missing features are complemented (e.g., FIG. 5, Embodiment 3 or 6).

[0174] In one embodiment, the step of complementing the missing features is performed when the multimodal features of the patient are at least 60% complete, at least 65% complete, at least 70% complete, at least 75% complete, or preferably at least 75% complete. The percentage of data completeness may be calculated relative to a complete set of features that can be extracted from the patient for the data.

[0175] Unless otherwise indicated, the term "at least" preceding a series of elements is to be understood as referring to all of the elements in the series. One of ordinary skill in the art can, using only routine experimentation, recognize or find many equivalents to the particular embodiments of the invention described herein. Such equivalents are intended to be encompassed by the present invention.

[0176] In one embodiment, a computer-implemented method is provided for predicting a treatment effect (predicting a treatment response or treatment efficacy) of a patient based on multimodal features of the patient, where in one step the features are aggregated into a feature value vector (also referred to as a feature vector value). In one embodiment, the multimodal features of the patient are not complete, and thus, this feature is aggregated into an incomplete feature value vector (an incomplete feature value vector). In one embodiment of the method of the present invention, which is performed prior to the step of complementing the multimodal features of a patient in which the calculation of at least one longitudinal feature is missing, clinical features, biological features, genomic features, and / or radiological features and at least one longitudinal feature are aggregated (e.g., FIG. 4, embodiments 2 or 5). Thus, in this embodiment, an incomplete feature value vector is obtained that includes or consists of clinical features, biological features, genomic features, and / or radiological features and at least one longitudinal feature (e.g., FIG. 4, embodiments 2 or 5).

[0177] In another embodiment, in the method of the present invention that is executed after the step of complementing the multimodal features of a patient lacking the calculation of at least one longitudinal feature, clinical features, biological features, genomic features, and / or radiological features are aggregated (e.g., FIG. 5, Embodiment 3 or 6). Thus, in one embodiment, an incomplete feature value vector including or consisting of clinical features, biological features, genomic features, and / or radiological features is obtained (e.g., FIG. 5, Embodiment 3 or 6). After the complementing step and after obtaining at least one longitudinal feature (included in the complete longitudinal feature value vector), a second aggregation step is executed, and clinical features, biological features, genomic features, and / or radiological features and at least one longitudinal feature are aggregated. In other words, the aggregation of the complete longitudinal feature value vector and the complete feature value vector is executed. Thus, in this embodiment, a completely aggregated multimodal longitudinal feature value vector including or consisting of clinical features, biological features, genomic features, and / or radiological features and at least one longitudinal feature is obtained (e.g., FIG. 5, Embodiment 3 or 6).

[0178] In one embodiment, a computer-implemented method for predicting the treatment response or treatment efficacy of a patient based on the multimodal features of the patient is provided. In one step, an incomplete feature value vector is the input to a trained complementary machine learning model, and a complete feature value vector is the output. Thus, the missing features are complemented by the trained complementary machine learning model.

[0179] In one embodiment, it is executed before the step of complementing the multimodal features of a patient lacking the calculation of at least one longitudinal feature. At this time, an incomplete feature value vector including or consisting of clinical features, biological features, genomic features, and / or radiological features and at least one longitudinal feature becomes the input to a trained complementary machine learning model (e.g., FIG. 4, Embodiment 2 or 5).

[0180] In another embodiment, the calculation of at least one longitudinal feature quantity is performed after the step of complementing the multi-modal feature quantity of a patient with missing values, where an incomplete feature quantity value vector including or consisting of clinical feature quantities, biological feature quantities, genomic feature quantities, and / or radiological feature quantities is input to a trained complementary machine learning model (e.g., FIG. 5, Embodiments 3 or 6).

[0181] In one particular embodiment, the missing feature quantity is complemented based on the same or different feature quantity modalities. In another particular embodiment, the missing feature quantity is complemented based on different feature quantity modalities (other feature quantity modalities), i.e., feature quantity modalities different from the feature quantity to be complemented, and the different feature quantity modalities are selected from the remaining three available feature quantity modalities. The feature quantity modalities (i.e., feature quantity types) can be selected from clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities. For example, the EGFR gene mutation status may be complemented from radiomics feature quantities, in which case the radiomics feature quantities are understood to be "other" or "different" feature quantity modalities from the genomic feature quantity modality.

[0182] In one embodiment, a computer-implemented method for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient) based on the multi-modal feature quantity of the patient is provided. In one step, a complete feature quantity value vector is filtered according to a list of highly informative feature quantity identifiers obtained in the training of the prediction machine learning model.

[0183] In one embodiment, a complete feature value vector or a completely aggregated multimodal longitudinal feature value vector may include or consist of clinical features, biological features, genomic features and / or radiological features and at least one longitudinal feature, and a list of highly informative feature identifiers may include highly informative clinical feature identifiers, highly informative biological feature identifiers, highly informative genomic feature identifiers, and / or highly informative radiological feature identifiers, and highly informative longitudinal feature identifiers (Embodiment 1 or 4).

[0184] Accordingly, a predicted feature value vector that is a subset of a complete feature value vector consisting of filtered features is obtained, and may be referred to herein as a predicted complete feature vector, a predicted feature value vector, a predicted feature vector value, a predicted complete feature vector or a predicted feature vector. In one embodiment, a predicted feature value vector that is a subset of a completely aggregated multimodal longitudinal feature value vector consisting of filtered features is obtained.

[0185] In one embodiment, a computer-implemented method for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient) based on the multimodal features of the patient is provided. In one step, this predicted feature value vector is an input to a trained predictive machine learning model, and the output is a prediction of the patient's response to this treatment, or the output is a prediction of the patient's treatment efficacy defined as the length of time to an event.

[0186] In one embodiment, the prediction of the treatment effect of the patient is classified as a prediction of the patient's response to the treatment. In one embodiment, the prediction of the patient's response to the treatment is classified as complete response, partial response, stable disease, progression. In another embodiment, the prediction of the patient's response to the treatment is classified as the likelihood of the patient's response to the treatment. In another embodiment, the prediction of the patient with respect to the treatment is binary classified as response or non-response.

[0187] In one embodiment, the prediction of the treatment effect of the patient is expressed as the prediction of the treatment efficacy of the patient. In one embodiment, the treatment efficacy of the patient is defined as the length of time to an event and is selected from progression-free survival (PFS), overall survival (OS), duration of response (DoR), and time to progression (TTP).

[0188] In another embodiment, a computer-implemented method for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient) is provided, the output of which is a list of informative feature identifiers used for training a predictive machine learning model, or a report weighted with the relative contribution of treatment features used in a method for predicting the treatment effect (treatment response or efficacy of the patient).

[0189] In another embodiment, a computer-implemented method for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient) is provided, where the prediction is made at a first evaluation time. Note that this prediction of the treatment effect (treatment response or treatment efficacy of the patient) is for a second evaluation time. These computer-implemented methods use as input the multimodal features of the patient collected at a baseline time point and / or the first evaluation time. This computer-implemented method uses the trained imputation machine learning model, the trained predictive machine learning model, and a list of informative feature identifiers, which are obtained using the features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is being made, collected at a baseline time point and / or the first evaluation time, and using the results collected at the first evaluation time and / or the second evaluation time.

[0190] In another embodiment, a computer-implemented method for predicting the treatment effect of a patient (predicting the treatment response or treatment efficacy of the patient) is provided, the prediction being made at a first evaluation time. Note that this prediction regarding the treatment response or treatment efficacy of the patient is for a second evaluation time.

[0191] In another embodiment, a computer-implemented method for predicting a patient's treatment effect (predicting a patient's treatment response or treatment efficacy) is provided, where the prediction is made as early as at a first evaluation time, but this prediction regarding the patient's treatment effect (prediction of the patient's treatment response or treatment efficacy) is for a next evaluation time, for example, the prediction is made at a second evaluation time for a third evaluation time, and can be made for such combinations.

[0192] A computer-implemented method for predicting a patient's treatment effect (predicting a patient's treatment response or treatment efficacy) at a second evaluation time uses, as input, the multimodal features of the patient collected at a baseline time point and / or at a first evaluation time. A computer-implemented method for predicting a patient's treatment effect (predicting a patient's treatment response) uses the trained complementary machine learning model, the trained prediction machine learning model, and a list of informative feature identifiers, which are obtained using the features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is being made, where the features are collected at a baseline time point and / or at a first evaluation time, and uses the results of treatment effectiveness (treatment response or treatment efficacy) at a first evaluation time and / or at a second evaluation time.

[0193] A computer-implemented method for predicting a patient's treatment effect at the earliest and second evaluation times (predicting a patient's treatment response or treatment efficacy) uses as input the multimodal features of the patient collected at the baseline time point and / or the first evaluation time. In one embodiment, a computer-implemented method for predicting a patient's treatment effect at the third evaluation time (predicting a patient's treatment response or treatment efficacy) uses as input the multimodal features of the patient collected at the baseline time point, and / or at the first evaluation time and / or at the second evaluation time. Thus, a computer-implemented method for predicting a patient's treatment effect (predicting a patient's treatment response or treatment efficacy) uses the trained complementary machine learning model, the trained prediction machine learning model, and the list of high-information-value feature identifiers, which are obtained using the features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is being made, the features being collected at the baseline time point, and / or at the first evaluation time and / or at the second evaluation time, and using the results of the treatment effect (treatment response or treatment efficacy) at the first evaluation time and / or at the second evaluation time and / or at the third evaluation time, and in such combinations.

[0194] It should be understood that not all features need to be collected at all time points. For example, if the displacement state is not likely to change, it may be sufficient to collect this feature only once.

[0195] In other words, a computer-implemented method for predicting a patient's treatment effect (predicting a patient's treatment response or efficacy) uses the trained complementary machine learning model, the trained prediction machine learning model, and the list of high-information-value feature identifiers, which are obtained / trained using the features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is being made, collected at the baseline time point, and the result of the treatment effect (treatment response or efficacy) was known at the first evaluation time, or the features were collected at the baseline time point and the first / further evaluation times, and the treatment effect was known at the second / further evaluation time.

[0196] In one embodiment, the selection of the multimodal features of the patient to be provided as input to a computer-implemented method for predicting the treatment effect (treatment response or treatment efficacy) of the patient is determined based on the condition of the specific patient.

[0197] In one embodiment, a set of the patient's multimodal data collected at the baseline time point and / or the first evaluation time point is preprocessed to be features, and the patient's treatment response (a patient treated with immunotherapy, chemotherapy, a combination of immunotherapy and chemotherapy, neoadjuvant therapy, targeted therapy, surgery, radiotherapy, hyperthermia, and / or adjuvant therapy for lung cancer) at the first evaluation time point and / or the second evaluation time point must be provided as input to a computer-implemented method for predicting the treatment response, where the patient's multimodal data includes: the treatment start date for the patient, the patient's response to treatment at the first evaluation time point, the date and progression index at the first evaluation time point, the date and survival index at the first evaluation time point, the PD-L1 expression level at the baseline time point, the radiological imaging data at the baseline time point and the first evaluation time point, the EGFR gene mutation status and the ALK gene mutation status at the baseline time point. Accordingly, the patient's multimodal features include: clinical features including the treatment start date for the patient, the response to treatment at the first evaluation time point, the date and progression index at the first evaluation time point, and the date and survival index at the first evaluation time point; biological features including the PD-L1 expression level at the baseline time point; radiomics features including features extracted from radiological imaging at the baseline time point and the first evaluation time point; and genomic features including the EGFR gene mutation status and the ALK gene mutation status at the baseline time point. In a further embodiment, the genomic features include the EGFR gene mutation status and the ALK gene mutation status collected at the baseline time point and / or the first evaluation time point.

[0198] In another embodiment, a computer-implemented method for predicting a patient's treatment effect (predicting treatment response or treatment efficacy) is provided, where the patient has cancer and the treatment is immunotherapy, chemotherapy, a combination of immunotherapy and chemotherapy, neoadjuvant therapy, targeted therapy, surgery, radiotherapy, hyperthermia, adjuvant postoperative therapy, and / or hormonal therapy.

[0199] In one embodiment, for patients with stage IV NSCLC treated with first-line treatment using pembrolizumab monotherapy, combination therapy of chemotherapy and pembrolizumab, or combination chemotherapy of two drugs, a computer-implemented method for predicting the treatment effect (predicting treatment response or treatment efficacy) of the patient for second-line treatment using immunotherapy such as anti-PD-1 / PD-L1, chemotherapy, targeted therapy, PARP inhibitors, or combinations thereof is provided.

[0200] In one embodiment, a computer-implemented method for predicting the treatment efficacy of a patient is provided, where the patient is predicted to respond completely or partially to the treatment, preferably predicted based on the computer-implemented method for predicting the treatment response of the patient described herein.

[0201] In one embodiment, a computer-implemented method for predicting a patient's treatment response or treatment efficacy is provided, the method comprising the following steps: a) the following, i. a trained imputation machine learning model trained to impute missing features of a patient, ii. a trained prediction machine learning model trained to predict a patient's treatment response or treatment efficacy, and iii. a list of high-information-value feature identifiers used for training the prediction machine learning model, wherein the step is to obtain, Using a set of multimodal features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and undergoing the same treatment, the complementary and predictive machine learning models are trained and the list of highly informative feature identifiers is obtained. For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, the obtaining step; b) receiving separately the multimodal features of the patient, the multimodal features of the patient including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, and the multimodal features of the patient being incomplete; at least one of the multimodal features of the patient is collected at at least two time points, the receiving step; c) aggregating the multimodal features of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step; d) inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector; e) filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; f) inputting the predicted feature value vector into the trained predictive machine learning model and outputting a prediction of the patient's response to the treatment or a prediction of the patient's treatment efficacy defined as the length of time to an event.

[0202] In another embodiment, a computer-implemented method for predicting a patient's treatment response or treatment efficacy is provided, the method including the following steps: a) the following, i. A trained completion machine learning model trained to complement the patient's missing feature quantities, ii. A trained prediction machine learning model trained to predict the patient's treatment response or treatment efficacy, and iii. A list of high-information-value feature identifiers used for training the prediction machine learning model, obtaining steps, Using at least two types of feature quantities selected from the clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities of a cohort of patients having the same disease as the patient for whom the prediction is to be made and receiving the same treatment, and / or at least one longitudinal feature quantity, the completion and prediction machine learning models are trained and the list of high-information-value feature identifiers is obtained, For each patient in the cohort, at least one of the multimodal feature quantities is collected at at least two time points, Metrics for changes between the values of at least one multimodal feature quantity of each patient collected at the at least two time points are calculated, For each patient, at least one longitudinal feature quantity is obtained, the obtaining step, b) Separately receiving the multimodal feature quantities of the patient including at least two types of feature quantities selected from clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities, the multimodal feature quantities of the patient being incomplete, At least one multimodal feature quantity of the patient is collected at at least two time points, Metrics for changes between the values of at least one received multimodal feature quantity of the patient collected at the at least two time points are calculated and at least one longitudinal feature quantity is obtained, The receiving step, which is performed before or after the step of complementing the feature quantities of the patient for which the calculation for at least one longitudinal feature quantity is missing, c) aggregating the multimodal features of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step, and d) inputting the feature value vector into the trained completion machine learning model to output a complete feature value vector, and e) filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values, and f) inputting the predicted feature value vector into the trained prediction machine learning model and outputting a prediction of the patient's response to the treatment or a prediction of the treatment efficacy of the patient defined as the length of time to an event.

[0203] In one embodiment, a computer-implemented method for predicting treatment response is provided, the method including the following steps: a) the following,[[]] i. a trained completion machine learning model trained to complement missing features of a patient, ii. a trained prediction machine learning model trained to predict a patient's treatment response, and iii. a list of highly informative feature identifiers used for training the prediction machine learning model, obtaining step, where using a set of multimodal features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and receiving the same treatment, the completion and prediction machine learning models are trained and the list of highly informative feature identifiers is obtained, for each patient in the cohort, at least one of the multimodal features is collected at at least two time points, the obtaining step, and b) receiving separately the multimodal features of the patient, including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein the multimodal features of the patient are not complete, the receiving step, wherein at least one multimodal feature of the patient is collected at at least two time points; c) aggregating the multimodal features of the patient into a feature value vector, wherein the feature value vector is not complete, the aggregating step; d) inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector; e) filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of the filtered feature values; f) inputting the predicted feature value vector into the trained prediction machine learning model and outputting a prediction of the patient's response to the treatment.

[0204] In one embodiment, a computer-implemented method for predicting the treatment efficacy of a patient is provided, the method comprising the following steps: a) the following, i. a trained complementary machine learning model trained to complement the missing features of a patient, ii. a trained prediction machine learning model trained to predict the treatment efficacy of a patient, and iii. a list of highly informative feature identifiers used for training the prediction machine learning model, the step of obtaining, Using a set of multimodal features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be performed and undergoing the same treatment, the complementary and predictive machine learning models are trained and the list of highly informative feature identifiers is obtained. For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, the obtaining step, b) separately receiving the multimodal features of the patient including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein the multimodal features of the patient are incomplete, at least one multimodal feature of the patient is collected at at least two time points, the receiving step, c) aggregating the multimodal features of the patient into a feature value vector, which feature value vector is incomplete, the aggregating step, d) inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector, e) filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values, f) inputting the predicted feature value vector into the trained predictive machine learning model and outputting a prediction of the treatment efficacy of the patient defined as the length of time to an event for the treatment.

[0205] In one embodiment, a computer-implemented method for predicting a patient's treatment response, referred to herein as Embodiment 1, is provided, the method comprising the following steps: a) the following, i. A trained completion machine learning model trained to complement the patient's missing feature quantities, ii. A trained prediction machine learning model trained to predict the patient's treatment response, and iii. A list of highly informative feature identifiers used for training the prediction machine learning model, obtaining, using at least two types of feature quantities selected from among the clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities of a cohort of patients having the same disease and undergoing the same treatment as the patient for whom the prediction is to be made, and / or at least one longitudinal feature quantity, the completion and prediction machine learning models are trained and the list of highly informative feature identifiers is obtained, for each patient in the cohort, at least one of the multimodal feature quantities is collected at at least two time points, metrics are calculated for the changes between the values of at least one multimodal feature quantity of each patient collected at the at least two time points, obtaining at least one longitudinal feature quantity for each patient, the obtaining step, b) receiving separately the multimodal feature quantities of the patient comprising at least two types of feature quantities selected from among clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities, the multimodal feature quantities of the patient being incomplete, at least one of the multimodal feature quantities of the patient is collected at at least two time points, metrics are calculated for the changes between the values of at least one received multimodal feature quantity of the patient collected at the at least two time points and at least one longitudinal feature quantity is obtained, the receiving step, performed before or after the step of complementing the (multimodal) feature quantities of the patient for which the calculation for at least one longitudinal feature quantity is missing, c) aggregating the multimodal features of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step, and d) inputting the feature value vector into the trained completion machine learning model to output a complete feature value vector; and e) filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of the filtered feature values; and f) inputting the predicted feature value vector into the trained prediction machine learning model and outputting a prediction of the patient's response to the treatment.

[0206] In another embodiment, a computer-implemented method for predicting a patient's treatment response, referred to herein as Embodiment 2 (Figure 4), is provided and includes the following steps: a) the following, i. a trained completion machine learning model trained to complement missing features of a patient, ii. a trained prediction machine learning model trained to predict a patient's treatment response, and iii. a list of highly informative feature identifiers used for training the prediction machine learning model, the obtaining step, using at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and receiving the same treatment, and at least one longitudinal feature, the completion and prediction machine learning models are trained and the list of highly informative feature identifiers is obtained, for each patient in the cohort, at least one of the multimodal features is collected at at least two time points, Metrics are calculated for changes between values of at least one multimodal feature of each patient collected at these at least two time points, obtaining, for each patient, at least one longitudinal feature, the obtaining step; b) receiving separately the multimodal features of the patient, including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein the multimodal features of the patient are not complete, receiving, wherein at least one multimodal feature of the patient is collected at at least two time points; c) calculating metrics for changes between values of at least one multimodal feature of the received patient collected at at least two time points to obtain at least one longitudinal feature of the patient; d) aggregating the multimodal features of the patient and at least one longitudinal feature of the patient into a feature value vector, wherein the feature value vector is not complete, the aggregating step; e) inputting the feature value vector into the trained completion machine learning model to output a complete feature value vector; f) filtering the plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; g) inputting the predicted feature value vector into the trained prediction machine learning model to output a prediction of the patient's response to the treatment.

[0207] In another embodiment, a computer-implemented method for predicting a patient's treatment response is provided, referred to herein as Embodiment 3 (Figure 5), and includes the following steps: a) the following i. a trained completion machine learning model trained to complete missing features of a patient, ii. A trained predictive machine learning model trained to predict a patient's treatment response, and iii. A list of highly informative feature identifiers used for training the predictive machine learning model, obtaining steps, At least two types of features selected from among the clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and receiving the same treatment, including a set of multimodal features and / or at least one longitudinal feature, using the complementary and predictive machine learning model is trained and the list of highly informative feature identifiers has been obtained, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, Metrics for the changes between the values of at least one multimodal feature of each patient collected at these at least two time points have been calculated, For each patient, at least one longitudinal feature has been obtained, the obtaining step, b) Separately receiving the multimodal features of the patient, including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein the multimodal features of the patient are incomplete, At least one of the multimodal features of the patient is collected at at least two time points, the receiving step, c) Aggregating the multimodal features of the patient into a feature value vector, wherein the feature value vector is incomplete, the aggregating step, d) Inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector, and e) Calculating a metric regarding the change between values of at least one multimodal feature of the received patient collected at at least two time points in the form of the complete feature value vector to obtain a complete longitudinal feature value vector; f) Aggregating the multimodal features of the patient in the complete feature value vector and at least one longitudinal feature of the patient in the complete longitudinal feature value vector to obtain a fully aggregated multimodal longitudinal feature value vector; g) Filtering a plurality of features of the fully aggregated multimodal longitudinal feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector which is a subset of the fully aggregated multimodal longitudinal feature value vector consisting of filtered feature values; h) Inputting the predicted feature value vector into the trained predictive machine learning model to output a prediction of the patient's response to the treatment.

[0208] In one embodiment, there is provided a computer-implemented method for predicting the treatment efficacy of a patient, which is referred to herein as Embodiment 4 and includes the following steps: a) The following, i. A trained imputation machine learning model trained to impute missing features of a patient, ii. A trained predictive machine learning model trained to predict the treatment efficacy of a patient, and iii. A list of highly informative feature identifiers used for training the predictive machine learning model, The step of obtaining, Using at least two types of features selected from the clinical features, biological features, genomic features and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be performed and receiving the same treatment, the imputation and predictive machine learning models are trained and the list of highly informative feature identifiers is obtained, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, metrics are calculated for the changes between the values of at least one multimodal feature of each patient collected at these at least two time points, the obtaining step, wherein at least one longitudinal feature is obtained for each patient, b) separately receiving the multimodal features of the patient, including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein the multimodal features of the patient are not complete, at least one multimodal feature of the patient is collected at at least two time points, metrics are calculated for the changes between the values of at least one received multimodal feature of the patient collected at these at least two time points, and at least one longitudinal feature is obtained, the receiving step, which is performed before or after the step of complementing the multimodal features of the patient for which the calculation of at least one longitudinal feature is missing, c) aggregating the multimodal features of the patient into a feature value vector, wherein this feature value vector is not complete, the aggregating step, d) inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector, e) filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values, f) inputting the predicted feature value vector into the trained prediction machine learning model and outputting a prediction of the treatment efficacy of the patient defined as the length of time until an event.

[0209] In one embodiment, there is provided a computer-implemented method for predicting the therapeutic efficacy of a patient, which is referred to herein as Embodiment 5 (Figure 4), and includes the following steps: a) The following, i. A trained completion machine learning model trained to complement the missing features of a patient, ii. A trained prediction machine learning model trained to predict the therapeutic efficacy of a patient, and iii. A list of high-information-value feature identifiers used for training the prediction machine learning model, which is a step of obtaining, Using at least two types of features selected from the clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and undergoing the same treatment, and at least one longitudinal feature, the completion and prediction machine learning models are trained and the list of high-information-value feature identifiers is obtained, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, Metrics are calculated for the changes between the values of at least one multimodal feature of each patient collected at the at least two time points, For each patient, at least one longitudinal feature is obtained, the obtaining step, and b) A step of separately receiving the multimodal features of the patient including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein the multimodal features of the patient are not complete, At least one multimodal feature of the patient is collected at at least two time points, the receiving step, and c) A step of calculating metrics for the changes between the values of at least one received multimodal feature of the patient collected at at least two time points to obtain at least one longitudinal feature of the patient, d) aggregating the multimodal features of the patient and at least one longitudinal feature of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step; e) inputting the feature value vector into the trained complementary machine learning model to output a complete feature value vector; f) filtering the plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; g) inputting the predicted feature value vector into the trained prediction machine learning model to output a prediction of the treatment efficacy of the patient defined as the length of time to an event.

[0210] In one embodiment, a computer-implemented method for predicting the treatment efficacy of a patient, referred to herein as embodiment 6 (Figure 5), is provided and includes the following steps: a) the following i. a trained complementary machine learning model trained to complement missing features of a patient, ii. a trained prediction machine learning model trained to predict the treatment efficacy of a patient, and iii. a list of highly informative feature identifiers used for training the prediction machine learning model, the step of obtaining, using at least two types of features selected from the clinical, biological, genomic, and radiological features of a cohort of patients having the same disease and undergoing the same treatment as the patient for whom the prediction is to be made, the complementary and prediction machine learning models are trained and the list of highly informative feature identifiers is obtained, for each patient in the cohort, at least one of the multimodal features is collected at at least two time points, Metrics are calculated for changes between values of at least one multimodal feature quantity of each patient collected at these at least two time points, obtaining at least one longitudinal feature quantity for each patient, the obtaining step, b) separately receiving the multimodal feature quantities of the patient, including at least two types of feature quantities selected from clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities, wherein the multimodal feature quantities of the patient are not complete, receiving, wherein at least one multimodal feature quantity of the patient is collected at at least two time points, c) aggregating the multimodal feature quantities of the patient into a feature quantity value vector, wherein this feature quantity value vector is not complete, the aggregating step, d) inputting the feature quantity value vector into the trained completion machine learning model to output a complete feature quantity value vector, e) calculating a metric for changes between values of at least one multimodal feature quantity of the received patient collected at at least two time points in the form of the complete feature quantity value vector to obtain a complete longitudinal feature quantity value vector, f) aggregating the multimodal feature quantities of the patient in the complete feature quantity value vector and at least one longitudinal feature quantity of the patient in the complete longitudinal feature quantity value vector to obtain a completely aggregated multimodal longitudinal feature quantity value vector, g) filtering a plurality of feature quantities of the completely aggregated multimodal longitudinal feature quantity value vector according to the list of highly informative feature quantity identifiers to obtain a predicted feature quantity value vector that is a subset of the completely aggregated multimodal longitudinal feature quantity value vector composed of filtered feature quantity values, h) inputting the predicted feature quantity value vector into the trained prediction machine learning model to output a prediction of the treatment efficacy of the patient defined as the length of time until an event.

[0211] Surprisingly, it has been demonstrated that the order of the steps of the methods according to Embodiments 2 and 5 above gives the best results. In an alternative embodiment, a computer-implemented method for predicting the treatment effect (predicting the treatment response or treatment efficacy) of a patient based on the patient's multimodal features is provided, where the features are complete. When the patient's multimodal features are complete, it should be understood that the computer-implemented method for predicting the treatment effect (predicting the treatment response or treatment efficacy) of the patient does not need to obtain a trained completion machine learning model trained to complement the patient's missing features, and does not perform the step of inputting the incomplete feature value vector into the trained completion machine learning model and outputting a complete feature value vector.

[0212] In one embodiment, a computer-implemented method for predicting a patient's treatment response is provided, the method comprising the following steps: a) The following, i. A trained prediction machine learning model trained to predict the treatment response of a patient, and ii. A list of highly informative feature identifiers used for training the prediction machine learning model, wherein the step of obtaining comprises: At least two types of features selected from among the clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and undergoing the same treatment, using a set of multimodal features and at least one longitudinal feature, the prediction machine learning model is trained and the list of highly informative feature identifiers is obtained, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, A metric for the change between the values of at least one multimodal feature of each patient collected at the at least two time points is calculated, For each patient, at least one longitudinal feature is obtained in the obtaining step; b) separately receiving the multi-modal features of the patient, including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, wherein the multi-modal features of the patient are complete; at least one multi-modal feature of the patient is collected at at least two time points; calculating a metric for the change between the values of at least one multi-modal feature of the received patient collected at the at least two time points to obtain at least one longitudinal feature; executing the calculating for at least one longitudinal feature before or after a feature aggregation step in the receiving step; c) aggregating the multi-modal features of the patient into a feature value vector, wherein the feature value vector is complete, in the aggregating step; d) filtering a plurality of features of the complete feature value vector according to the list of informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; e) inputting the predicted feature value vector into the trained predictive machine learning model and outputting a prediction of the patient's response to the treatment.

[0213] In a further embodiment, a computer-implemented method for predicting the treatment efficacy of a patient is provided, the method including the steps of: a) the following: i. a trained predictive machine learning model trained to predict the treatment efficacy of a patient, and ii. a list of informative feature identifiers used for training the predictive machine learning model, in the obtaining step, Using a set of multimodal features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease as the patient for whom the prediction is to be made and undergoing the same treatment, the predictive machine learning model is trained and the list of highly informative feature identifiers is obtained. For each patient in the cohort, at least one of the multimodal features is collected at at least two time points. Metrics are calculated for the changes between the values of at least one multimodal feature of each patient collected at the at least two time points. The step of obtaining, for each patient, at least one longitudinal feature, b) The step of separately receiving the multimodal features of the patient, the multimodal features of the patient including at least two types of features selected from clinical features, biological features, genomic features, and radiological features, the multimodal features of the patient being complete. At least one of the multimodal features of the patient is collected at at least two time points. Metrics are calculated for the changes between the values of at least one received multimodal feature of the patient collected at the at least two time points, and at least one longitudinal feature is obtained. The step of receiving, wherein the calculation for at least one longitudinal feature is performed before or after the feature aggregation step. c) The step of aggregating the multimodal features of the patient into a feature value vector, the feature value vector being complete, the step of aggregating. d) The step of filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predictive feature value vector that is a subset of the complete feature value vector consisting of filtered feature values. e) Inputting the predicted feature value vector into the trained predictive machine learning model and outputting a prediction of the therapeutic efficacy of the patient, defined as the length of time until a certain event.

[0214] In one further alternative embodiment, a computer-implemented method for predicting the treatment effect of a patient is provided, the method comprising the steps of: a) The following, i. A trained imputation machine learning model trained to impute missing features of a patient, ii. A trained predictive machine learning model trained to predict the treatment effect of a patient, and iii. A list of informative feature identifiers used for training the predictive machine learning model, wherein the step of obtaining comprises: Using at least two types of features selected from the clinical, biological, genomic, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made, the imputation and predictive machine learning models are trained and the list of informative feature identifiers is obtained, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, A metric for the change between the values of at least one multimodal feature of each patient collected at the at least two time points is calculated, At least one longitudinal feature is obtained for each patient, the obtaining step; and b) Separately receiving the multimodal features of the patient, comprising at least two types of features selected from clinical, biological, genomic, and radiological features, wherein the multimodal features of the patient are incomplete, the receiving step; c) aggregating the multimodal feature quantities of the patient into a feature quantity value vector, the feature quantity value vector being incomplete, the aggregating step, and d) inputting the feature quantity value vector into the trained completion machine learning model to output a complete feature quantity value vector; and e) filtering a plurality of feature quantities of the complete feature quantity value vector according to the list of highly informative feature quantity identifiers to obtain a predicted feature quantity value vector that is a subset of the complete feature quantity value vector consisting of filtered feature quantity values; and f) inputting the predicted feature quantity value vector into the trained prediction machine learning model and outputting a prediction of the treatment effect of the patient. Accordingly, a computer-implemented method according to an embodiment of the present invention is provided, the prediction being made at a baseline time point for a first evaluation time, the multimodal feature quantities of the patient being collected at the baseline time point, after which the completion and prediction machine learning models are trained and a list of highly informative feature quantity identifiers is obtained, but the multimodal feature quantities of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is being made, the multimodal feature quantities collected at the baseline time point and the first evaluation time, and optionally using the results of the treatment response at the first evaluation time, the list having been obtained.

[0215] In one embodiment, a data processing apparatus is provided that includes means for performing the method described herein. In one embodiment, a computer program is provided that includes instructions that, when the program is executed by a computer, cause the computer to perform the method described herein.

[0216] In one embodiment, a computer-readable medium is provided that includes instructions that, when executed by a computer, cause the computer to perform the method described herein.

[0217] In one embodiment, when the computer-implemented method described herein executes the steps described herein such that the computer-implemented method of the present invention consists of the method steps described herein, it provides a technical result of predicting the treatment effect of a patient.

[0218] Patient and Treatment In one embodiment, a patient according to the present invention suffers from a neurological disease, cancer, genetic heart disease or genetic neurological disease, or a rare genetic disease such as Pompe disease.

[0219] In a particular embodiment, a patient according to the present invention suffers from cancer. In another particular embodiment, a patient according to the present invention suffers from a brain tumor, breast cancer such as triple-negative breast cancer (TNBC), kidney cancer, head and neck cancer, ovarian cancer or colon cancer.

[0220] In another particular embodiment, a patient according to the present invention suffers from lung cancer, particularly non-small cell lung cancer (NSCLC). In one embodiment, a patient according to the present invention is receiving treatment for a neurological disease, genetic heart disease or genetic neurological disease, or a rare genetic disease such as Pompe disease.

[0221] In another particular embodiment, a patient according to the present invention is receiving treatment for cancer, particularly immunotherapy, chemotherapy (such as neoadjuvant chemotherapy (NCT)), targeted therapy, treatment using anti-angiogenic agents, surgery, radiotherapy or a combination of these treatments.

[0222] In another particular embodiment, a patient according to the present invention is receiving treatment for cancer, particularly immunotherapy, chemotherapy, a combination of immunotherapy and chemotherapy, targeted therapy, surgery, radiotherapy, hyperthermia, and / or hormone therapy or a combination of these treatments in the context of neoadjuvant therapy, adjuvant therapy or maintenance therapy.

[0223] In one embodiment, immunotherapy includes treatment using an immune checkpoint inhibitor anti-PD-(L)1 antibody or anti-CTLA-4 antibody, such as pembrolizumab, nivolumab, atezolizumab, durvalumab, semaprilizumab, dostarlimab, and the like.

[0224] In one embodiment, the patient is diagnosed with stage IV NSCLC (new or early stage progression to stage IV). In another specific embodiment, the patient has stage IV non-small cell lung cancer (NSCLC) and is diagnosed with NSCLC that has no cancer gene driver mutations in the EGFR gene or the ALK gene that can be targeted by treatment in first-line treatment. Patients with stage IV NSCLC can be treated with pembrolizumab monotherapy, combination therapy of chemotherapy and pembrolizumab, or combination chemotherapy with two drugs.

[0225] In one embodiment, the patient has stage IV NSCLC and is diagnosed with NSCLC that has no activating mutations in cancer genes eligible for targeted therapy. Example 1. Multimodal machine learning model for predicting response to ICI treatment in patients with metastatic NSCLC Patients diagnosed with stage IV non-small cell lung cancer (NSCLC) without cancer gene driver mutations in the EGFR gene or the ALK gene that can be targeted by treatment have limited treatment options. Some of these patients respond to ICI in first-line treatment. However, current biomarkers for identifying which patients are likely to respond to treatment on a long-term scale are limited.

[0226] It was devised to develop a multimodal method for predicting patient response to treatment for a cohort of 65 stage IV NSCLC patients treated by combination therapy of chemotherapy and pembrolizumab. Various clinical data, biological data, histological data, radiological data, and CT-scan imaging data were collected for each patient up to the start of treatment. CT scan imaging was also collected at the first evaluation occurring between 2 and 3 months.

[0227] A total of 240 radiomics features were extracted from 3D lung tumor slices obtained from CT scans acquired at the baseline time point and after the first evaluation time point, and combined with other data modalities to predict the long-term progression status of each patient after the start of treatment.

[0228] First, a data mining process is applied that includes recording of the categorized features and exclusion of features with more than 50% missing values. Next, hierarchical clustering is applied to reduce the dimensionality of the radiomics features collected before the start of treatment and after the first evaluation. Using the bootstrap method to evaluate the stability of the obtained partitions, 14 clusters were obtained for the radiomics features. Each cluster is manually characterized based on the univariate relationship between the feature and its outcome, and the interpretability of that feature. The rate of change for the radiomics features between the baseline and the first evaluation time point is calculated by computer. These are examples of preprocessing steps for the data. The features used for the machine learning algorithm are the baseline features, clinical and biological variables finally collected at the first evaluation time point, the baseline radiomics features, and the rate of change for the radiomics features between the baseline and the first evaluation time point.

[0229] The best ML model achieved an iAUC (integrated AUC) of 0.86 over a period between 2 and 20 months. This included baseline data, first evaluation data and the rate of change between the baseline and the first evaluation time point.

Claims

1. A computer-implemented method for predicting the treatment effect of a patient, the method comprising: a) the following: i. A trained completion machine learning model trained to complement the missing feature values of a patient; ii. A trained prediction machine learning model trained to predict the treatment effect of a patient; and iii. A list of high-information-value feature identifiers used for training the prediction machine learning model, wherein the step of obtaining comprises: using at least two types of feature values selected from the clinical, biological, genomic, and radiological feature values of a cohort of patients having the same disease as the patient for whom the prediction is to be made and receiving the same treatment, and at least one longitudinal feature value, to train the completion and prediction machine learning models and obtain the list of high-information-value feature identifiers, for each patient in the cohort, at least one of the multimodal feature values is collected at at least two time points, metrics are calculated for the changes between the values of at least one multimodal feature value of each patient collected at the at least two time points, obtaining at least one longitudinal feature value for each patient, the obtaining step; and b) receiving separately the multimodal feature values of the patient, the multimodal feature values of the patient including at least two types of feature values selected from clinical, biological, genomic, and radiological feature values, and the multimodal feature values of the patient being incomplete, at least one of the multimodal feature values of the patient is collected at at least two time points, metrics are calculated for the changes between the values of at least one received multimodal feature value of the patient collected at the at least two time points, and at least one longitudinal feature value is obtained, the receiving step being performed before or after the step of complementing the multimodal feature values of the patient for which the calculation for at least one longitudinal feature value is missing, c) aggregating the multimodal feature values of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step; and d) inputting the feature value vector into the trained completion machine learning model to output a complete feature value vector. e) Filtering a plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predictive feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; f) Inputting the predictive feature value vector into the trained predictive machine learning model and outputting a prediction of the treatment effect of the patient; A method comprising. **Claim 2** The computer-implemented method for predicting the treatment effect of a patient according to Claim 1, Performed before the step of complementing the multimodal features of a patient in which the calculation for the at least one longitudinal feature is missing, the method comprising: a) The following, i. A trained completion machine learning model trained to complement missing features of a patient, ii. A trained predictive machine learning model trained to predict the treatment effect of a patient, and iii. A list of highly informative feature identifiers used for training the predictive machine learning model, The step of obtaining, Using at least two types of features selected from the clinical, biological, genomic, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made, the completion and predictive machine learning models are trained and the list of highly informative feature identifiers is obtained, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, Metrics have been calculated for the changes between the values of at least one multimodal feature of each patient collected at the at least two time points, For each patient, at least one longitudinal feature has been obtained, the step of obtaining; b) Separately receiving the multimodal features of the patient, comprising at least two types of features selected from clinical, biological, genomic, and radiological features, wherein the multimodal features of the patient are not complete, At least one of the multimodal features of the patient is collected at at least two time points, the step of receiving; c) calculating a metric for a change between values of at least one multimodal feature of the received patient collected at at least two time points to obtain at least one longitudinal feature of the patient; d) aggregating the multimodal features of the patient and the at least one longitudinal feature of the patient into a feature value vector, the feature value vector being incomplete, the aggregating step; e) inputting the feature value vector into the trained completion machine learning model to output a complete feature value vector; f) filtering the plurality of features of the complete feature value vector according to the list of highly informative feature identifiers to obtain a predicted feature value vector that is a subset of the complete feature value vector consisting of filtered feature values; g) inputting the predicted feature value vector into the trained prediction machine learning model to output a prediction of the treatment effect of the patient; A method comprising. **Claim 3** The computer-implemented method for predicting a treatment effect of a patient according to claim 1, wherein the calculation for the at least one longitudinal feature is performed after a step of complementing missing multimodal features of the patient, the method comprising: a) the following i. A trained completion machine learning model trained to complement missing features of a patient, ii. A trained prediction machine learning model trained to predict a treatment effect of a patient, and iii. A list of highly informative feature identifiers used for training the prediction machine learning model, wherein the step of obtaining, using at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be performed, the completion and prediction machine learning models are trained and the list of highly informative feature identifiers is obtained, for each patient in the cohort, at least one of the multimodal features is collected at at least two time points, and a metric for a change between values of at least one multimodal feature of each patient collected at the at least two time points is calculated. For each patient, at least one longitudinal feature quantity has been acquired, said acquiring step; b) receiving separately the multimodal feature quantities of said patient, said multimodal feature quantities of said patient including at least two types of feature quantities selected from clinical feature quantities, biological feature quantities, genomic feature quantities and radiological feature quantities, and said multimodal feature quantities of said patient being incomplete; said receiving step, wherein at least one multimodal feature quantity of said patient is collected at at least two time points; c) aggregating the multimodal feature quantities of said patient into a feature quantity value vector, said aggregating step, and said feature quantity value vector being incomplete; d) inputting said feature quantity value vector into said trained completion machine learning model to output a complete feature quantity value vector; e) calculating a metric for the change between the values of at least one multimodal feature quantity of said received patient collected at at least two time points in the form of said complete feature quantity value vector to obtain a complete longitudinal feature quantity value vector; f) aggregating the multimodal feature quantities of said patient in said complete feature quantity value vector and at least one longitudinal feature quantity of said patient in said complete longitudinal feature quantity value vector to obtain a completely aggregated multimodal longitudinal feature quantity value vector; g) filtering a plurality of feature quantities of said completely aggregated multimodal longitudinal feature quantity value vector according to said list of highly informative feature quantity identifiers to obtain a predicted feature quantity value vector which is a subset of the completely aggregated multimodal longitudinal feature quantity value vector consisting of filtered feature quantity values; h) inputting said predicted feature quantity value vector into said trained prediction machine learning model to output a prediction of the treatment effect of said patient; A method comprising.

4. The computer-implemented method according to any one of claims 1 to 3, wherein said prediction of the treatment effect of said patient is represented as a prediction of the response of said patient to said treatment, said trained prediction machine learning model being trained to predict the treatment effect of a patient represented as the response of said patient to said treatment.

5. The computer-implemented method according to claim 4, wherein said prediction of the response of said patient to said treatment is as complete response, partial response, stable disease or progression, A method classified as the likelihood of the patient's response to said treatment.

6. The computer-implemented method according to any one of claims 1 to 3, wherein said prediction of the treatment effect of said patient is represented as a prediction of the treatment efficacy of said patient, wherein said trained predictive machine learning model is trained to predict the treatment effect of said patient represented as the treatment efficacy of said patient defined as the length of time to an event.

7. The computer-implemented method according to claim 6, wherein the treatment efficacy of said patient is defined as the length of time to an event and is selected from progression-free survival (PFS), overall survival (OS), duration of response (DoR), and time to progression (TTP).

8. The computer-implemented method according to any one of claims 1 to 7, wherein said prediction is made for a second evaluation time at a first evaluation time, wherein the multimodal features of said patient are collected at a baseline time point and at a first evaluation time, using the multimodal features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom said prediction is performed, collected at a baseline time point and at a first evaluation time, and using the results of the treatment response at a second evaluation time, the complementary and predictive machine learning model is trained, and the list of informative feature identifiers is obtained.

9. The computer-implemented method according to any one of claims 1 to 5, wherein said patient has cancer and said treatment is immunotherapy, chemotherapy (e.g., neoadjuvant chemotherapy (NCT), etc.), targeted therapy, treatment using an anti-angiogenic agent, surgery, radiation therapy, or a combination of these treatments.

10. The computer-implemented method according to claim 9, wherein said patient has lung cancer and said treatment is immunotherapy, chemotherapy, a combination of immunotherapy and chemotherapy, neoadjuvant therapy, targeted therapy, treatment using an anti-angiogenic drug, surgery, radiation therapy, hyperthermia, and / or adjuvant therapy after surgery, wherein the multimodal features of said patient are the treatment start date for said patient, the response to treatment at a first evaluation, the date and an indicator of progression at a first evaluation, the date and an indicator of survival at a first evaluation, including a plurality of clinical features, and a plurality of biological features including the expression level of PD-L1 at a baseline time point. A plurality of radiomics features including a plurality of feature quantities extracted from radiological imaging data at the baseline time point and in the first evaluation, A plurality of genomic features including the EGFR gene mutation state and the ALK gene mutation state at the baseline time point, A method comprising:

11. The computer-implemented method according to any one of claims 1 to 7, The training step of the complementary machine learning model inputs a set of multimodal features including at least two types of feature quantities selected from clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made into a machine learning learned by a training algorithm, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, The trained complementary machine learning model generates, as an output, a complete list of a plurality of feature quantities for a certain patient from an incomplete list. A method.

12. The computer-implemented method according to claim 11, Metrics regarding changes between values of at least one multimodal feature collected at at least two time points for each patient in the cohort are further calculated, at least one longitudinal feature quantity is obtained for each patient in the cohort, and The training step of the complementary machine learning model further includes inputting the at least one longitudinal feature quantity into a machine learning learned by a training algorithm. A method.

13. The computer-implemented method according to claim 4, The training step of the predictive machine learning inputs a set of multimodal features including at least two types of feature quantities selected from clinical feature quantities, biological feature quantities, genomic feature quantities, and radiological feature quantities of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made and at least one longitudinal feature quantity into a machine learning learned by a training algorithm, For each patient in the cohort, at least one of the multimodal features is collected at at least two time points, Metrics regarding changes between values of at least one multimodal feature of each patient collected at the at least two time points are calculated, At least one longitudinal feature is obtained for each patient, and the trained predictive machine learning model outputs a classification of the patient's response to the treatment or a likelihood of the patient's response to the treatment and a list of high-information-value feature identifiers used for training the predictive machine learning model. A method.

14. The computer-implemented method according to claim 6, wherein the training step of the predictive machine learning inputs into the machine learning, which is learned by a training algorithm, a set of features including at least two types of features selected from clinical features, biological features, genomic features, and radiological features of a cohort of patients having the same disease and receiving the same treatment as the patient for whom the prediction is to be made, and at least one longitudinal feature, for each patient in the cohort, at least one of the multimodal features is collected at at least two time points, a metric is calculated for the change between the values of at least one multimodal feature of each patient collected at the at least two time points, at least one longitudinal feature is obtained for each patient, and the trained predictive machine learning model outputs a classification of the treatment efficacy defined as the length of time until an event and a list of high-information-value feature identifiers used for training the predictive machine learning model. A method.

15. The computer-implemented method according to any one of claims 1 to 7, wherein the output is supplemented by the list of high-information-value feature identifiers used for training the predictive machine learning model and / or a report with a list regarding the relative contribution of a plurality of features used in the method for predicting, for example, the treatment response or treatment efficacy of a patient, the treatment effect of the patient. A method.

16. The computer-implemented method according to any one of claims 1 to 7, wherein features are complemented based on different feature modalities. A method.

17. The computer-implemented method according to any one of claims 1 to 7, wherein the multimodal features of the patient are at least 75% complete. A method.

Images