System and method for predicting probability of ecog performance status of lung cancer patients using interpretable deep learning and program stored in a recording medium

Abstract

According to one embodiment of the present invention, a system for predicting the functional state of a lung cancer patient may be provided, comprising: a sample data collection unit for collecting sample data including data on a plurality of explanatory variables and outcome variables for predicting the functional state of a lung cancer patient; a variable selection unit for selecting explanatory variables to be used in a learning model for predicting the functional state of a lung cancer patient among a plurality of explanatory variables of the sample data; a learning model construction unit for performing learning on a plurality of learning models based on the selected explanatory variables and outcome variables of the collected sample data and constructing a plurality of learning models for predicting the functional state of a lung cancer patient; a learning model selection unit for selecting a learning model for predicting the functional state of a lung cancer patient based on performance statistics from the constructed plurality of learning models; an explanatory model generation unit for generating a local surrogate model for explaining the correlation between the functional state of a lung cancer patient and explanatory variables based on the selected learning model; and a prediction result providing unit for providing a functional state prediction result that predicts the functional state of a lung cancer patient, who is the subject of analysis, from the explanatory variables of the subject of analysis based on the selected learning model and the generated local surrogate model.

Description

Technology Field

[0001] The present invention relates to a technology for predicting the functional state of a lung cancer patient, and more specifically, to a system for predicting the functional state of a lung cancer patient using interpretable deep learning, a method for predicting the functional state of a lung cancer patient, and a program stored on a recording medium. Background Technology

[0002] Lung cancer is not only one of the most widespread cancers globally but also the leading cause of cancer-related deaths. In Korea, the 5-year survival rate for terminal lung cancer patients was 8.9% in 2019, making it a type of cancer with a poor prognosis. However, the possibility of a complete cure (5-year survival rate) is increasing due to the recent expansion of health screenings.

[0003] The functional status of lung cancer patients is a crucial factor not only for predicting prognosis but also for determining quality of life. Functional status is closely linked to the onset of cancer symptoms, implying that functional capacity deteriorates as symptoms worsen. To extend the survival period and improve the quality of life for lung cancer patients, it is important to screen their functional status early.

[0004] The Eastern Cooperative Oncology Group (ECOG) status is a widely used method for evaluating the functional capabilities of cancer patients and predicting their prognosis. Various factors, including age and psychological symptoms, can influence the ECOG status of lung cancer patients. ECOG status provides important information to healthcare professionals, as it reflects the patient's health condition and can change over time. Therefore, it is crucial for healthcare providers to frequently assess the ECOG status of lung cancer patients to accurately reflect their current functional capabilities and modify treatment plans accordingly. The problem to be solved

[0005] The present invention was developed to solve the problems of the prior art as described above. The problem to be solved by the present invention is to provide a system for predicting the functional state of a lung cancer patient using interpretable deep learning, a method for predicting the functional state of a lung cancer patient, and a program that predicts the functional state of a lung cancer patient with high accuracy while allowing medical personnel to understand the prediction results. means of solving the problem

[0006] As a means for solving the above-mentioned problem, a system for predicting the functional state of a lung cancer patient according to one embodiment of the present invention may include: a sample data collection unit that collects sample data including data on a plurality of explanatory variables and result variables for predicting the functional state of a lung cancer patient; a variable selection unit that selects explanatory variables to be used in a learning model for predicting the functional state of a lung cancer patient among a plurality of explanatory variables of the sample data; a learning model construction unit that performs learning on a plurality of learning models based on the selected explanatory variables and result variables of the collected sample data and constructs a plurality of learning models for predicting the functional state of a lung cancer patient; a learning model selection unit that selects a learning model for predicting the functional state of a lung cancer patient based on performance statistics from the constructed plurality of learning models; an explanatory model generation unit that generates a local surrogate model for explaining the correlation between the functional state of a lung cancer patient and the explanatory variables based on the selected learning model; and a prediction result providing unit that provides a functional state prediction result for predicting the functional state of a lung cancer patient, who is the subject of analysis, from the explanatory variables of the subject of analysis based on the selected learning model and the generated local surrogate model.

[0007] The explanatory variables to be used in a learning model to predict the functional status of lung cancer patients can be selected based on the Extra Tree (ET) learning model.

[0008] The selected multiple explanatory variables may include (1) presence or absence of a thoracoscopic / thoracoscopic surgical approach, (2) presence or absence of surgery, (3) presence or absence of a thoracotomy / open surgical approach, (4) type of adenocarcinoma, (5) presence or absence of clinical stage 1 of trilobate lymph nodes (#10), (6) presence or absence of supportive care, (7) ALK FISH test results, (8) presence or absence of clinical stage 2 of subtracheal lymph nodes (#4), (9) presence or absence of symptoms of dyspnea, (10) duration of smoking, (11) presence or absence of second-line chemotherapy, (12) presence or absence of radiation therapy, and (13) presence or absence of third-line chemotherapy.

[0009] Multiple learning models constructed to predict the functional status of lung cancer patients may include an AdaBoost model and a TabNet model. The selected learning model may include an AdaBoost model.

[0010] The above local surrogate model can calculate the contribution rates of multiple explanatory variables to the functional status of lung cancer patients in order to explain the correlation between the functional status of lung cancer patients and explanatory variables.

[0011] Listing the contribution rates of multiple explanatory variables to the functional status of lung cancer patients from largest to smallest, the order may be (1) presence or absence of dyspnea symptoms, (2) presence or absence of subtracheal lymph nodes (#4) clinical stage 2, (3) presence or absence of thoracoscopic / thoracoscopic surgical approach, (4) presence or absence of surgery, (5) duration of smoking, (6) presence or absence of supportive care, (7) type of adenocarcinoma, (8) presence or absence of third-line chemotherapy, (9) presence or absence of radiation therapy, (10) presence or absence of thoracotomy / open surgical approach, (11) presence or absence of trilobate lymph nodes (#10) clinical stage 1, (12) presence or absence of second-line chemotherapy, and (13) ALK FISH test results.

[0012] Preprocessing can be performed to adjust the sample data into balanced data by classifying samples in the above sample data with an ECOG status scale in the range of 1 to 5 as “physical activity limited” and samples with an ECOG status scale of 0 as “no functional limitation”.

[0013] As another means for solving the above-mentioned problem, a method for predicting the functional state of a lung cancer patient according to an embodiment of the present invention may include: a sample data collection step of collecting sample data including data on a plurality of explanatory variables and outcome variables for predicting the functional state of a lung cancer patient; a variable selection step of selecting explanatory variables to be used in a learning model for predicting the functional state of a lung cancer patient among a plurality of explanatory variables of the sample data; a learning model construction step of performing learning on a plurality of learning models based on the selected explanatory variables and outcome variables of the collected sample data and constructing a plurality of learning models for predicting the functional state of a lung cancer patient; a learning model selection step of selecting a learning model for predicting the functional state of a lung cancer patient based on performance statistics from the constructed plurality of learning models; an explanatory model generation step of generating a local surrogate model for explaining the correlation between the functional state of a lung cancer patient and explanatory variables based on the selected learning model; and a prediction result provision step of providing a functional state prediction result that predicts the functional state of a lung cancer patient, who is the subject of analysis, from the explanatory variables of the subject of analysis based on the selected learning model and the generated local surrogate model.

[0014] As another means for solving the above-mentioned problem, a program according to one embodiment of the present invention may include a program stored on a recording medium to perform a method for predicting the functional state of a lung cancer patient by a computer. Effects of the invention

[0015] According to one embodiment of the present invention, a system for predicting the functional state of a lung cancer patient, a method for predicting the functional state of a lung cancer patient using interpretable deep learning, and a program may be provided, which predict the functional state of a lung cancer patient with high accuracy while allowing medical personnel to understand the prediction results. This is achieved by collecting sample data including data on a plurality of explanatory variables and result variables for predicting the functional state of a lung cancer patient, selecting explanatory variables to be used in a learning model for predicting the functional state of a lung cancer patient from among the plurality of explanatory variables of the sample data, performing learning on a plurality of learning models based on the selected explanatory variables and result variables of the collected sample data, constructing a plurality of learning models for predicting the functional state of a lung cancer patient, selecting a learning model for predicting the functional state of a lung cancer patient based on performance statistics from the constructed plurality of learning models, generating a local surrogate model for explaining the correlation between the functional state of a lung cancer patient and explanatory variables based on the selected learning model, and providing a functional state prediction result that predicts the functional state of a lung cancer patient from the explanatory variables of the subject of analysis based on the selected learning model and the generated local surrogate model. Brief explanation of the drawing

[0016] FIG. 1 is a block diagram illustrating the configuration of a system for predicting the functional state of a lung cancer patient according to one embodiment of the present invention. FIG. 2 is a diagram showing the configuration of a system for predicting the functional state of a lung cancer patient according to one embodiment of the present invention when implemented on a computer. FIG. 3 is a diagram showing the configuration of a system for predicting the functional state of a lung cancer patient according to one embodiment of the present invention, implemented by a server and a client terminal. FIG. 4 is a flowchart of a method for predicting a risk group for depressive disorder implemented by a functional state prediction system for lung cancer patients according to one embodiment of the present invention. Figure 5 is a diagram illustrating an overview of the pipeline of a prediction model according to the present invention. Figure 6 is a diagram explaining the concept of AdaBoost. Figure 7 is a diagram explaining the concept of TabNet. Figure 8 is a diagram showing the interpretation for a specific lung cancer patient in a local surrogate model. Figure 9 is a graph showing the contribution to the prediction of ECOG functional status. Specific details for implementing the invention

[0017] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

[0018] The embodiments of the present invention are provided to more fully explain the invention to those skilled in the art, and the following embodiments may be modified in various different forms, and the scope of the invention is not limited to the following embodiments. Rather, these embodiments are provided to make the disclosure more faithful and complete and to fully convey the spirit of the invention to those skilled in the art.

[0019] The various embodiments described herein may be implemented, for example, in a recording medium readable by a computer or similar device using software, hardware, or a combination thereof.

[0020] According to hardware implementation, the embodiments described herein may be implemented using at least one of ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), processors, controllers, microcontrollers, microprocessors, and electrical units for performing functions.

[0021] According to the software implementation, embodiments such as procedures or functions may be implemented together with separate software modules that perform at least one function or operation. The software code may be implemented by a software application written in a suitable programming language.

[0022] FIG. 1 is a block diagram illustrating the configuration of a functional state prediction system (10) for a lung cancer patient according to one embodiment of the present invention.

[0023] Referring to FIG. 1, a functional state prediction system (10) for lung cancer patients may include a sample data collection unit (110), a variable selection unit (120), a learning model construction unit (130), a learning model selection unit (140), an explanation model generation unit (150), and a prediction result provision unit (160). The functional state prediction system (10) for lung cancer patients may be implemented by a processor and memory installed in an information processing terminal. Additionally, the functional state prediction system (10) for lung cancer patients may be implemented by a processing server connected to the information processing terminal via a network.

[0024] The components of the functional status prediction system (10) for lung cancer patients are described in detail below, and these components may be implemented in the form of modules in an information processing terminal or in the form of services provided by a server and a client.

[0025] The sample data collection unit (110) can collect sample data including data on multiple explanatory variables and outcome variables for predicting the functional status of a lung cancer patient. Sample data on multiple explanatory variables and outcome variables can be obtained from the lung cancer patient's hospital record data.

[0026] Multiple explanatory variables of the sample data may include data such as gender, smoking history, age, body mass index, histopathological type, symptoms, clinical stage, ECOG score, treatment method, and survival status. To prevent overfitting to the sample data, balancing preprocessing of the sample data may be performed.

[0027] The variable selection unit (120) can select explanatory variables to be used in a learning model for predicting the functional status of lung cancer patients from among a plurality of explanatory variables of sample data. Among a plurality of explanatory variables including gender, smoking history, age, body mass index, histopathological type, symptoms, clinical stage, ECOG score, treatment method, and survival status, explanatory variables having an effect on the functional status of lung cancer patients greater than or equal to a predetermined value may be selected as explanatory variables to be used in the learning model.

[0028] The selection of explanatory variables is useful for identifying and ranking important and highly valued characteristics in sample data. In one embodiment, explanatory variables (features) can be extracted using Extra Tree (ET) activation.

[0029] Multiple explanatory variables selected using ET (Extra Tree) activation may include (1) presence or absence of thoracoscopic / thoracoscopic surgical approach, (2) presence or absence of surgery, (3) presence or absence of thoracotomy / open surgical approach, (4) type of adenocarcinoma, (5) presence or absence of trilobate lymph node (#10) clinical stage 1, (6) presence or absence of supportive care, (7) ALK FISH test results, (8) presence or absence of subtracheal lymph node (#4) clinical stage 2, (9) presence or absence of dyspnea symptoms, (10) duration of smoking, (11) presence or absence of second-line chemotherapy, (12) presence or absence of radiation therapy, and (13) presence or absence of third-line chemotherapy.

[0030] The learning model building unit (130) performs learning on multiple learning models based on selected explanatory variables and result variables of collected sample data, and can build multiple learning models to predict the functional state of lung cancer patients. The multiple learning models built to predict the functional state of lung cancer patients may include an AdaBoost model and a TabNet model. Subsequently, among the multiple learning models, one model having a higher performance statistic based on the prediction performance statistic may be selected.

[0031] The learning model selection unit (140) can select a learning model for predicting the functional state of a lung cancer patient based on performance statistics from a plurality of constructed learning models. In one embodiment, the selected learning model may be an AdaBoost model. The performance of the plurality of learning models and the performance of the selected AdaBoost model will be described later.

[0032] The explanatory model generation unit (150) can generate a local surrogate (LIME) model to explain the correlation between the functional state of a lung cancer patient and explanatory variables based on a selected learning model. The local surrogate model can calculate the contribution rate of multiple explanatory variables to the functional state of a lung cancer patient in order to explain the correlation between the functional state of a lung cancer patient and explanatory variables.

[0033] Neural network models are black-box models because they consist of a very deep, multi-layered structure in which nodes in each layer are complexly connected; consequently, it is impossible to directly interpret which elements of the input data had a significant impact on the result. The LIME algorithm is an algorithm that performs a local surrogate analysis method, which interprets each data point by reconstructing it into a human-interpretable representation rather than interpreting the entire dataset.

[0034] Given input data x, the LIME algorithm aims to find an interpretable model g∈G that predicts the result closest to a neural network model f from a set of interpretable models G. To perform the LIME algorithm, an interpretable representation vector x' and sample data z' are first created using it. Then, the sample data z' and data z (converted into an original format to be used as input for the model to be interpreted) are used as inputs to interpretable model g and neural network model f, respectively, to obtain their respective predictions. Finally, the model g that most closely approximates model f is found by utilizing the difference between these predictions.

[0035] [Equation 1]

[0036]

[0037] [Equation 1] is the loss function L and the complexity for each interpretable model g It is a formula for finding an interpretable model g that produces the result closest to the result of the neural network model f using , and is the similarity between the data x to be explained and the sampled data z.

[0038] The explanation model generation unit (150) can be constructed by steps including: (1) a first step of checking the prediction results of a basic learning model and selecting observations related to the patient to be analyzed; (2) a second step of obtaining prediction data by inputting modified input data obtained by modifying the explanatory variables of the sample data into the basic learning model; (3) a third step of assigning weights to the modified input data obtained in the second step according to the proximity to observations related to the patient to be analyzed; and (4) a fourth step of training a local surrogate model with weights obtained in the third step by inputting the modified input data obtained in the second step into the basic learning model as an explanatory variable and using the prediction data of the basic learning model obtained as a result variable.

[0039] Listing the contribution rates of multiple explanatory variables to the functional status of lung cancer patients from largest to smallest, the order may be (1) presence or absence of dyspnea symptoms, (2) presence or absence of subtracheal lymph nodes (#4) clinical stage 2, (3) presence or absence of thoracoscopic / thoracoscopic surgical approach, (4) presence or absence of surgery, (5) duration of smoking, (6) presence or absence of supportive care, (7) type of adenocarcinoma, (8) presence or absence of third-line chemotherapy, (9) presence or absence of radiation therapy, (10) presence or absence of thoracotomy / open surgical approach, (11) presence or absence of trilobate lymph nodes (#10) clinical stage 1, (12) presence or absence of second-line chemotherapy, and (13) ALK FISH test results.

[0040] The prediction result providing unit (160) can provide a functional state prediction result that predicts the functional state of the lung cancer patient being analyzed from the explanatory variables of the subject being analyzed, based on the selected learning model and the generated local surrogate model.

[0041] Input variable data is obtained from the patient's questionnaire and input into the basic learning model and the LIME learning model. Consequently, the basic learning model outputs a predicted value of the patient's functional status, while the LIME learning model outputs the contribution rate for each explanatory variable. Medical professionals can predict the patient's prognosis and determine an appropriate treatment method by referring to the outputted predicted value of functional status and the contribution rate for each explanatory variable.

[0043] FIG. 2 is a diagram showing the configuration of a system (10) for predicting the functional state of a lung cancer patient according to one embodiment of the present invention when implemented on a computer.

[0044] The computer device (200) in which the functional status prediction system (10) for lung cancer patients is implemented refers to an information processing device such as a PC, laptop, smart device, or server. The computer device (100) may include an input device (210), a computational device (220), a storage device (230), and an output device (240).

[0045] The input device (210) can receive multiple sample data and measurement values ​​for explanatory variables of the prediction subject. The input device (210) can receive sample data in the form of a data file or receive sample data via an input device such as a keyboard or mouse. The received sample data and measurement values ​​for explanatory variables of the prediction subject can be stored in a storage device (230).

[0046] Additionally, the input device (210) can receive the user's voice using a device such as a microphone and use text converted from the voice as input data. According to the voice input method, in an emergency environment such as an emergency room, medical personnel can use text data converted from voice input in a learning model without typing the input data of an emergency patient.

[0047] The storage device (230) can store input sample data and measurements of explanatory variables of the prediction target. Additionally, the storage device (230) can store the constructed basic learning model and LIME learning model.

[0048] The computing device (220) can build a basic learning model for predicting survival probability and a LIME learning model for calculating contribution rates by learning sample data through machine learning. Additionally, the computing device (220) can calculate a predicted survival probability value of a predicted subject and a contribution rate for each explanatory variable from the built learning model.

[0049] The output device (240) is a device that outputs survival probability and contribution rates by explanatory variable in a specific form. The output device (140) may include at least one of a display device, an output device that outputs a document, and a communication device that transmits survival probability / contribution rate information to another device.

[0051] FIG. 3 is a diagram showing the configuration of a lung cancer patient functional state prediction system (10) according to one embodiment of the present invention, implemented by a server (330) and a client terminal (310).

[0052] The functional status prediction system (10) for lung cancer patients includes a client terminal (210), a test DB (220), a processing server (230), and a model DB (240).

[0053] The client terminal (210) may be installed in a place such as a public institution, a medical institution, or a patient's home. The client terminal (210) may receive input of sample data or measurements of explanatory variables of the prediction subject. The input sample data or measurements of explanatory variables of the prediction subject may be stored in the inspection DB (220).

[0054] The inspection DB (220) may include a database that stores input sample data and measurements of explanatory variables of the prediction subject.

[0055] The processing server (330) can build a learning model to learn sample data and calculate the predicted value of the functional status of lung cancer patients and the contribution rate for each explanatory variable. Additionally, the processing server (330) can calculate the predicted value of the functional status and contribution rate of the predicted subject from the built learning model.

[0056] The model DB (340) may contain data related to the constructed basic learning model and LIME learning model. The processing server (330) may store the constructed basic learning model and LIME learning model in the model DB (340).

[0057] Although the inspection DB (320), processing server (330), and model DB (340) are shown as separate components in FIG. 3, the inspection DB (320), processing server (330), and model DB (340) can be configured as a single integrated component.

[0059] FIG. 4 is a flowchart of a method for predicting the functional state of a lung cancer patient implemented by a system (10) for predicting the functional state of a lung cancer patient according to one embodiment of the present invention.

[0060] Referring to FIG. 4, a method for predicting the functional status of a lung cancer patient may include a sample data collection step (S110), a variable selection step (S120), a learning model construction step (S130), a learning model selection step (S140), an explanation model generation step (S150), and a prediction result provision step (S160).

[0061] In the sample data collection step (S110), the sample data collection unit (110) may collect sample data including data on multiple explanatory variables and result variables for predicting the functional status of a lung cancer patient.

[0062] In the variable selection step (S120), the variable selection unit (120) can select an explanatory variable to be used in a learning model for predicting the functional state of a lung cancer patient from among a plurality of explanatory variables of the sample data.

[0063] In the learning model building step (S130), the learning model building unit (130) performs learning on a plurality of learning models based on selected explanatory variables and result variables of the collected sample data, and can build a plurality of learning models to predict the functional state of lung cancer patients.

[0064] In the learning model selection step (S140), the learning model selection unit (140) can select a learning model to predict the functional state of a lung cancer patient based on performance statistics from a plurality of established learning models.

[0065] In the explanation model generation step (S150), the explanation model generation unit (150) can generate a local surrogate model to explain the correlation between the functional state of a lung cancer patient and the explanation variable based on the selected learning model.

[0066] In the prediction result providing step (S160), the prediction result providing unit (160) may provide a functional state prediction result that predicts the functional state of the lung cancer patient being analyzed from the explanatory variables of the subject being analyzed, based on the selected learning model and the generated local surrogate model.

[0067] With respect to the sample data collection step (S110), variable selection step (S120), learning model construction step (S130), learning model selection step (S140), explanation model generation step (S150), and prediction result provision step (S160), the above description of the sample data collection unit (110), variable selection unit (120), learning model construction unit (130), learning model selection unit (140), explanation model generation unit (150), and prediction result provision unit (160) may be referenced.

[0069] The following describes the features of the functional status prediction system and prediction method for lung cancer patients according to the present invention through an explanation of the development and verification process of the functional status prediction model for lung cancer patients.

[0071] [Data Sources and Preprocessing]

[0072] The experimental data consists of 2,063 domestic lung cancer patients from the National Cancer Center in 2016. Data collected included gender, smoking history, age, body mass index, histopathological type, symptoms, clinical stage, ECOG score, treatment method, and survival status.

[0073] Preprocessing of the dataset was required before developing a model with the experimental data. The collected raw data contained a total of 2,829 healthy patients and 328 identification information columns. Initially, 18 duplicate columns related to serial numbers and date / time values ​​were removed. Next, columns containing 50% null values ​​were removed from the dataset to account for the visualized values. Finally, 2,063 patients with 286 variables and no missing ECOG variables were selected. The dataset was split into training and testing sets using an 80:20 split to complete the essential setup for the model.

[0074] The sample size of each score value in the ECOG status scale and dataset is presented in Table 1. Referring to Table 1, there is a possibility that machine model training may be overfitted during the prediction phase because score 0 is excessively saturated compared to other scores. Therefore, in this technique, without changing the size of the dataset, score values ​​from 1 to 5 were re-categorized as "physical activity limited" and score 0 as "no functional limitation." The ratio between the number of score 0 values ​​(Class 0) and score 1 values ​​(Class 1) was 904:1159, confirming that the data is balanced.

[0076] [Table 1]

[0077]

[0078] [Select Feature]

[0079] Feature selection is important before developing machine model training. Feature selection methods help identify the most important and highly valued features in a dataset and rank them first. The inventors extracted features using Extra Tree (ET) activation. Table 2 shows the features selected by ET activation.

[0080] [Table 2]

[0081]

[0082] [Development and Validation of a Learning Model for Predicting the Functional Status of Lung Cancer Patients]

[0083] The prediction model of the present invention was developed using the Python programming language. Figure 5 illustrates an overview of the pipeline of the prediction model according to the present invention. The inventors evaluated the machine learning model using a stratified 10-fold cross-validation (CV) method. When utilizing k-fold CV, the dataset is divided into k equal-sized sections, and cases are randomly selected for each fold. A portion of each subset is used for testing, and the remainder is used as the training set. Each subset is utilized as the test set once during the k evaluations of the model. During stratified k-fold cross-validation, each subset is divided into groups with class labels that are nearly identical to those of the original dataset.

[0084] The core concept of the AdaBoost algorithm is to construct a unique tactical friction of learnable patterns (set training) based on previous results at each step (for each classifier). Figure 6 is a diagram illustrating the concept of AdaBoost.

[0085] The inventors fine-tuned the model's hyperparameters using the Optuna library, which automatically searches for a framework of hyperparameters.

[0086] TabNet is a deep learning model suitable for structured data built on sequential multi-step processing. TabNet improves the ability to learn multi-dimensional features and assists in feature selection. Each n-th step processes a D-dimensional feature vector and generates an output for the Feature Block Transformer.

[0087] Figure 7 illustrates the concept of TabNet. The attentive transformer, mask, feature transformer, split node, and ReLu activation are all included in each step. The model's hyperparameters were fine-tuned using the Optuna framework.

[0088] [Performance Evaluation of the Training Model]

[0089] After implementing the model, the performance of the proposed machine learning model was evaluated using accuracy, precision, recall, and F1 score.

[0090] The calculations for Accuracy, Precision, Recall, and F1-score, which are metrics for evaluating the performance of the DNN in this technology, are as follows.

[0091]

[0092]

[0093]

[0094]

[0095] Additionally, the inventors evaluated the performance of the model using ROC AUC. While AUC summarizes the model's total diagnostic tests, the ROC curve represents the TPR (positive rate) and FPR (false positive rate) of the diagnostic tests. The range of the ROC AUC metric is [0, 1], where 0 indicates a complete result, 0.7–0.8 is considered a good grade, and 0.8 or higher is considered to indicate very good predictive performance. The definition of ROC AUC is as follows.

[0096]

[0097] The inventors assumed that the model with the largest ROC AUC is the model with the best predictive performance, and if the ROC AUC is the same, defined the model with the higher recall as the superior model.

[0098] [Development of Explainable Machine Learning Using Local Surrogate Models]

[0099] Explaining the results of a machine learning system is summarized as establishing a connection between the input to the system and the results generated by the system at a level understandable to humans. Modern machine learning systems, such as deep learning-based predictive models, involve highly sophisticated parameter designs behind the model's conclusions that make the inference difficult to understand. In this invention, a local surrogate model is used to develop an explanation for the model output representing the correlation between the functional status of lung cancer patients and other variables.

[0100] Local surrogate modeling techniques utilize interpretable local models to explain each unique prediction. The algorithm assigns weights to additional data points in relation to the original point, and ultimately, the dataset uses a surrogate model, such as linear regression, utilizing the sample weights.

[0101] Each raw data point utilizes a trained explanatory model, Ω(g: a measure of complexity). In classification, f(x) represents the probability (or binary indicator) that x belongs to a specific class, and (z) is additionally used as a measure of proximity between instance z and x to determine locality around x. Next, it measures how inaccurately g approximates f in a locale denoted by πx(z). At this point, the algorithm minimizes Ω(g) to ensure local fidelity and interpretability, keeping it low enough for humans to interpret. Through this procedure, the explanation of the local surrogate model is finally obtained.

[0102] [Performance of the TabNet Model]

[0103] To deploy an ideal model using 2,063 patients and 13 major variables, the optimal parameters of the AdaBoost and TabNet models were optimized by Optuna using the superparameters of the AdaBoost and TabNet models. The fine-tuning results derived from this model are as follows.

[0104]

[0105] The AdaBoost model outperformed the TabNet model on the set data. Specifically, the AdaBoost model had an ROC AUC of 0.7735 and a recall of 0.8284, while the TabNet model had an ROC AUC of 0.7395 and a recall of 0.6898. The performance scores of the models are presented in Table 3. The observable AdaBoost model received higher scores than the TabNet model across all performance metrics. Additionally, the AdaBoost model had the highest recall.

[0106] [Table 3]

[0107]

[0108] [LIME-based AdaBoost Model Evaluation]

[0109] The inventors selected a specific case to analyze in order to demonstrate how the LIME model works with the AdaBoost model to predict the functional status (ECOG) of cancer patients after treatment.

[0110] Figure 8(c) describes a lung cancer patient with limited physical activity, summarizing the patient's condition and related circumstances. The patient in this case had never undergone surgery (OP_NO = 0) and had not been treated with thoracoscopic surgery / VATS (OP_APLACH_2 = 0) or open thoracotomy / open-lung surgical approaches (OP_APLACH_1 = 0). Additionally, this patient did not have adenocarcinoma (S_NAME_2 = 0), had lower tracheal lymph nodes (#4) damaged at clinical stage 2 (HC_NSITE2_3 = 0), had not received second or third-line chemotherapy (MED2_LINE = 0, MED3_LINE = 0), had received radiation therapy (RT = 0), had no symptoms of dyspnea (SYMPTS_4 = 0), and did not require supportive care (TX_COM4 = 0). As a result of inputting this patient into the prediction model, it was predicted that the patient's physical function (ECOG) would be limited with an 84% probability, as shown in Figure 8 (a).

[0111] Figure 8(b) visually presents the LIME-AdaBoost model in this technology. As a result of this case demonstration, “thoracoscopic and VATS surgical approach (OP_APPROACH_2), stage 2 lower paratracheal lymph nodes (#4) (HC_NSITE2_3), number of operations (OP_NO), adenocarcinoma (S_NAME_2), second-line chemotherapy (MED2_LINE), open thoracotomy / open-lung surgical approach (OP_APPROACH_1), radiation therapy (RT), and third-line chemotherapy (MED3_LINE)” were identified as the major factors and conditions that contribute most to the prediction of physical function in lung cancer patients.

[0112] Figure 9 is a graph showing the contribution to the prediction of ECOG functional status.

[0114] The steps or processes described above may be executed by hardware components, software components, and / or a combination of hardware components and software components. For example, the steps or processes described in the embodiments may be executed using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0115] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or command the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0116] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

[0117] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0118] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A system for predicting the functional state of a lung cancer patient comprising: a sample data collection unit for collecting sample data including data on multiple explanatory variables and outcome variables for predicting the functional state of a lung cancer patient; a variable selection unit for selecting explanatory variables to be used in a learning model for predicting the functional state of a lung cancer patient among multiple explanatory variables of the sample data; a learning model construction unit for performing learning on multiple learning models based on the selected explanatory variables and outcome variables of the collected sample data and constructing multiple learning models for predicting the functional state of a lung cancer patient; a learning model selection unit for selecting a learning model for predicting the functional state of a lung cancer patient based on performance statistics from the constructed multiple learning models; an explanatory model generation unit for generating a local surrogate model to explain the correlation between the functional state of a lung cancer patient and explanatory variables based on the selected learning model; and a prediction result providing unit for providing a functional state prediction result that predicts the functional state of a lung cancer patient, who is the subject of analysis, from the explanatory variables of the subject of analysis based on the selected learning model and the generated local surrogate model. Claim 2 A system for predicting the functional state of a lung cancer patient according to claim 1, wherein the explanatory variables to be used in a learning model for predicting the functional state of a lung cancer patient are selected based on an Extra Tree (ET) learning model. Claim 3 A system for predicting the functional status of a lung cancer patient according to claim 1, wherein the selected multiple explanatory variables include (1) presence or absence of a thoracoscopic / thoracoscopic surgical approach, (2) presence or absence of surgery, (3) presence or absence of a thoracotomy / open surgical approach, (4) type of adenocarcinoma, (5) presence or absence of clinical stage 1 of trilobate lymph nodes (#10), (6) presence or absence of supportive care, (7) ALK FISH test results, (8) presence or absence of clinical stage 2 of subtracheal lymph nodes (#4), (9) presence or absence of symptoms of dyspnea, (10) duration of smoking, (11) presence or absence of second-line chemotherapy, (12) presence or absence of radiation therapy, and (13) presence or absence of third-line chemotherapy. Claim 4 In claim 1, a plurality of learning models constructed to predict the functional status of a lung cancer patient include an AdaBoost model and a TabNet model, a lung cancer patient functional status prediction system. Claim 5 In claim 1, the selected learning model is a functional status prediction system for lung cancer patients including an adaBoost model. Claim 6 In claim 1, the local surrogate model is a system for predicting the functional state of a lung cancer patient that calculates the contribution rate of a plurality of explanatory variables to the functional state of a lung cancer patient in order to explain the correlation between the functional state of a lung cancer patient and explanatory variables. Claim 7 A system for predicting the functional status of lung cancer patients according to claim 1, wherein the magnitudes of the contribution rates of multiple explanatory variables to the functional status of lung cancer patients are listed from largest to smallest as follows: (1) presence or absence of dyspnea symptoms, (2) presence or absence of subtracheal lymph nodes (#4) clinical stage 2, (3) presence or absence of thoracoscopic / thoracoscopic surgical approach, (4) presence or absence of surgery, (5) duration of smoking, (6) presence or absence of supportive care, (7) type of adenocarcinoma, (8) presence or absence of third-line chemotherapy, (9) presence or absence of radiation therapy, (10) presence or absence of thoracotomy / open surgical approach, (11) presence or absence of trilobate lymph nodes (#10) clinical stage 1, (12) presence or absence of second-line chemotherapy, and (13) ALK FISH test results. Claim 8 A system for predicting the functional status of lung cancer patients according to claim 1, wherein preprocessing is performed to adjust the sample data into balanced data by classifying samples in the sample data with an ECOG status scale in the range of 1 to 5 as "physical activity limited" and samples with an ECOG status scale of 0 as "no functional limitation". Claim 9 A method for predicting the functional state of a lung cancer patient, comprising: a sample data collection step of collecting sample data including data on multiple explanatory variables and outcome variables for predicting the functional state of a lung cancer patient; a variable selection step of selecting explanatory variables to be used in a learning model for predicting the functional state of a lung cancer patient from among the multiple explanatory variables of the sample data; a learning model construction step of performing training on multiple learning models based on the selected explanatory variables and outcome variables of the collected sample data and constructing multiple learning models for predicting the functional state of a lung cancer patient; a learning model selection step of selecting a learning model for predicting the functional state of a lung cancer patient based on performance statistics from the constructed multiple learning models; an explanatory model generation step of generating a local surrogate model to explain the correlation between the functional state of a lung cancer patient and the explanatory variables based on the selected learning model; and a prediction result provision step of providing a functional state prediction result that predicts the functional state of a lung cancer patient, who is the subject of analysis, from the explanatory variables of the subject of analysis based on the selected learning model and the generated local surrogate model. Claim 10 A program stored on a recording medium to perform the method for predicting the functional status of a lung cancer patient described in paragraph 9 by a computer.

Images