Method for determining treatment for cancer patient

Abstract

To better predict a patient's risk of early death and better inform appropriate treatment for individual clinical condition.SOLUTION: A method is disclosed for determining a treatment for a cancer patient, comprising: measuring concentrations of a plurality of clinicopathological markers for gradient boosting machine learning analysis; and receiving a prediction as to whether a patient has a likelihood of dying within a 12-week period, from the gradient boosting machine learning model. The method further comprises administering an anti-cancer therapeutic agent to a patient when a prediction indicates that the patient does not have the likelihood of dying within the 12-week period.SELECTED DRAWING: None

Description

[Technical field]

[0001] The present invention relates generally to methods of treating patients for cancer. [Background technology]

[0002] The development of new anti-cancer therapeutics, such as cancer immunotherapy (IO) and drugs targeting the DNA damage repair system, requires the investment of significant time and resources to identify potentially effective drug candidates. However, before such candidates can be added to the existing set of approved anti-cancer treatments, they must be proven both safe and effective in clinical trials. The first step in this process is to conduct Phase I clinical trials, which are designed to focus on drug safety.

[0003] Such Phase I clinical trials often recruit patients who have failed standard treatments and therefore may benefit from participating in the trial. In many clinical trials, regardless of the type of treatment being tested or its intended indication, a subset of participants will not be able to complete the trial for a variety of reasons. For example, there may be patients who participate in a clinical trial with significantly advanced underlying disease or who have conditions secondary to their cancer or other unrelated conditions.

[0004] In the context of IO Phase I clinical trials, such as PD-1 or PD-L1 (PDx) trials as monotherapy, nearly 20% of recruited patients die within 2-3 months of study initiation, even though most of these trials ideally last much longer than 2-3 months. Therefore, to participate in IO Phase I clinical trials, patients must have a life expectancy of more than 24 weeks and be at no risk of early death (EM; defined as a life expectancy of less than 24 weeks). In many randomized clinical trials comparing ICIs with active control groups in patients with advanced or metastatic cancer, the EM phenomenon has been observed, even if an overall benefit in favor of ICI therapy was ultimately found (Champiat, et al. 2018). The exact cause of this phenomenon has not been clearly established, but it is characterized by a disproportionately higher mortality rate in the control group during the early treatment period, when survival is favorable, followed by a favorable overall survival (OS) benefit in the ICI-treated group. In clinical data, this is often reflected by the “Kaplan-Meier curve crossing” that suggests a subpopulation of patients at high risk for EM whose rate of tumor growth progression may require the cytotoxic cytoreductive effects of chemotherapy.

[0005] Therefore, it is important in treatment to better predict the risk of early death for a given patient to better inform the appropriate treatment for their individual clinical condition. If a large number of patients participating in a Phase I clinical trial experience unexpected rapid disease progression or worsening of another underlying pathology, this may reduce the ability to detect early efficacy signals of the test treatment, delay dose escalation due to early dropout of patients, confound the analysis of predictive biomarkers of efficacy, and dilute efficacy against controls in randomized trials. This may also prevent patients from receiving therapeutic interventions that may extend life outside of Phase I clinical trials. If such patients were to receive a therapeutic intervention that corrects or alleviates the underlying pathology, they would be able to fully benefit from their participation in the trial. Therefore, there is a need for improved prognostic methodologies to predict EM in patients undergoing anticancer therapeutic clinical trials.

[0006] Immune checkpoint inhibitors (ICIs) have revolutionized the treatment of many cancer types, including melanoma, non-small cell lung cancer (NSCLC), renal cell carcinoma (RCC), squamous cell carcinoma of the head and neck (SCCHN), urothelial carcinoma (UC), and Hodgkin lymphoma, and are associated with long-lasting tumor responses. To serve this purpose, we present herein a novel prognostic score, the Immune Immediacy Index (3i score), that uses pretreatment measurements of routinely collected blood-based factors to predict a patient's risk of early death and optimize the benefit (risk profile of treatment for ICI patients). Summary of the Invention [Means for solving the problem]

[0007] As described below, in a first aspect, the disclosure provides a method of treating a patient having cancer, the method comprising: a) determining concentrations of a plurality of clinicopathological markers for the patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; c) receiving a prediction from the gradient boosting machine learning model of whether the patient has a likelihood of dying within a 12-week period; and d) administering an anti-cancer therapeutic to the patient if the prediction indicates that the patient does not have a likelihood of dying within a 12-week period.

[0008] In one embodiment of the first aspect, the plurality of clinicopathological markers are age, albumin (g / L), alanine phosphatase (U / L), alanine aminotransferase (U / L), aspartate aminotransferase (U / L), basophils (10 3 / μL), basophils / white blood cells (%), bilirubin (μmol / L), body mass index, calcium (mmol / L), chloride (mmol / L), eosinophils (10 3 / μL), eosinophils / leukocytes (%), gamma-glutamyltransferase (U / L), glucose (mmol / L), hematocrit (%), hemoglobin (g / L), potassium (mmol / L), lactate dehydrogenase (U / L), liver metastasis (yes / no), lymphocytes (103 / μL), lymphocytes / white blood cells (%), magnesium (mmol / L), monocytes (10 3 / μL), monocytes / leukocytes (%), neutrophils (10 3 / μL), neutrophils / leukocytes (%), neutrophils / lymphocytes (%), platelets (10 3 / μL), protein (g / L), gender, sodium (mmol / L), thyrotropin (mU / L), tumor size, uric acid (mmol / L) and white blood cells (10 3 / μL).

[0009] In another embodiment of the first aspect, the plurality of clinicopathological markers include neutrophils / white blood cells (%), neutrophils (10 3 / μL), albumin (g / L), lactate dehydrogenase (U / L), aspartate aminotransferase (U / L) and / or gamma-glutamyltransferase (U / L).

[0010] In one embodiment of the first aspect, the method further comprises obtaining a sample from the patient.

[0011] In one embodiment of the first aspect, the sample comprises a tissue biopsy, cerebrospinal fluid, lymphatic fluid, whole blood, serum, blood cells, urine, sweat, tears, saliva and / or faeces.

[0012] In one embodiment of the first aspect, the concentration of one or more of the plurality of clinical pathological markers is determined from a sample taken from the patient.

[0013] In one embodiment of the first aspect, the anticancer therapeutic comprises cisplatin, gemcitabine, methotrexate, vinblastine, doxorubicin, cisplatin (MVAC), carboplatin, taxanes, temozolomide, dacarbazine, vinflunine, docetaxel, paclitaxel, nab-paclitaxel, vemurafenib, erlotinib, afatinib, cetuximab, bevacizumab, gefitinib, and / or pemetrexed.

[0014] In one embodiment of the first aspect, the anti-cancer therapeutic comprises an immune checkpoint inhibitor.

[0015] In one embodiment of the first aspect, the immune checkpoint inhibitor comprises an anti-CTLA-4 antibody, an anti-PD-1 antibody and / or an anti-PD-L1 antibody.

[0016] In one embodiment of the first aspect, the anti-CTLA-4 antibody is tremelimumab or ipilimumab.

[0017] In one embodiment of the first aspect, the anti-PD-1 antibody is REGN2810, SHR1210, IBI308, PDR001, nivolumab, pembrolizumab, anti-PD-1, BGB-A317, BCD-100, or JS001.

[0018] In one embodiment of the first aspect, the anti-PD-L1 antibody comprises durvalumab, avelumab, atezolizumab or KNO35.

[0019] In a further embodiment of the first aspect, the method includes determining a type of cancer of the patient and also providing the determined type of cancer to a gradient boosting machine learning model.

[0020] In a second aspect, the disclosure provides a method of treating a patient having cancer, the method comprising: a) determining concentrations of a plurality of clinicopathological markers for the patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; c) receiving a prediction from the gradient boosting machine learning model of whether the patient has reduced immune fitness; and d) administering an anti-cancer therapeutic to the patient if the prediction indicates that the patient does not have reduced immune fitness.

[0021] In one embodiment of the second aspect, the method further comprises e) reducing in the patient at least one of the rate of cancer cell division, tumor growth, tumor size, tumor density, or rate of tumor metastasis.

[0022] In a third aspect, the disclosure provides a method of treating a patient having cancer, the method comprising: a) determining concentrations of a plurality of clinicopathological markers for the patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; c) receiving a prediction from the gradient boosting machine learning model of whether the patient has poor immune fitness; d) administering a therapeutic agent to the patient to increase the patient's immune fitness if the prediction indicates that the patient has poor immune fitness; e) determining second concentrations of the plurality of clinicopathological markers for the patient; f) providing the determined second concentrations of the plurality of clinicopathological markers to the gradient boosting machine learning model; g) receiving a second prediction from the gradient boosting machine learning model of whether the patient has poor immune fitness; and h) administering an anti-cancer therapeutic agent to the patient if the prediction indicates that the patient does not have poor immune fitness.

[0023] In a fourth aspect, the disclosure provides a method of conducting a clinical trial of an anti-cancer therapeutic, the method comprising: a) identifying a candidate participant for the clinical trial, the candidate participant having cancer; b) determining concentrations of a plurality of clinicopathological markers for the patient; c) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; d) receiving a prediction from the gradient boosting machine learning model of whether the patient has a likelihood of dying within a 12 week period; and e) including the patient in the clinical trial of the anti-cancer therapeutic if the prediction indicates that the patient does not have a likelihood of dying within a 12 week period.

[0024] In one embodiment of the fourth aspect, the anti-cancer therapeutic is an immune checkpoint inhibitor.

[0025] In a fifth aspect, the present disclosure provides a method of predicting efficacy of treating a cancer patient with an immune checkpoint inhibitor, comprising: a) determining concentrations of a plurality of clinicopathological markers for the patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; and c) receiving a prediction from the gradient boosting machine learning model of whether the patient lacks immune fitness, wherein a prediction indicating that the patient does not lack immune fitness predicts that the patient may be treated more effectively with an immune checkpoint inhibitor (ICI) than a patient predicted to lack immune fitness.

[0026] In a sixth aspect, the present disclosure provides a method of determining a therapeutic regimen for a cancer patient, the method comprising: a) determining concentrations of a plurality of clinicopathological markers for the patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; and c) receiving a prediction from the gradient boosting machine learning model of whether the patient is deficient in immune fitness, wherein a prediction indicating the patient is not deficient in immune fitness suggests a therapeutic regimen including an immune checkpoint inhibitor (ICI), and a prediction indicating the patient is deficient in immune fitness suggests a therapeutic regimen that does not include an immune checkpoint inhibitor (ICI).

[0027] In a seventh aspect, a computing system includes a computer memory configured to store model training data, the model training data including a first set of cancer patients undergoing cancer treatment, the model training data associating (i) results from laboratory tests performed on the first set of cancer patients with tumor types of the first set of cancer patients, and (ii) whether each patient from the first set of cancer patients died within a reference number of weeks from the start of the cancer treatment. The computing system also includes one or more processors in communication with the computer memory and configured to execute program instructions for: (i) training a series of decision trees based on the model training data until a stopping condition is reached, each successive decision tree in the series of decision trees being constructed based on a residual value of its predecessor in the series of decision trees; and (ii) creating a gradient boosting machine learning model as an additive function of the series of decision trees, the gradient boosting machine learning model predicting whether a further cancer patient will die within a reference number of weeks based on the additive function applied to the results from laboratory tests performed on the further cancer patients and the tumor types of the further cancer patients.

[0028] In an eighth aspect, the computer-implemented method includes obtaining model training data, the model training data including a first set of cancer patients who have undergone cancer treatment, the model training data associating (i) results from a clinical test performed on the first set of cancer patients with tumor types of the first set of cancer patients, and (ii) whether each patient from the first set of cancer patients died within a reference number of weeks from the start of the cancer treatment. The computer-implemented method includes training a series of decision trees based on the model training data until a stopping condition is reached, each subsequent decision tree in the series of decision trees being constructed based on a residual value of its predecessor in the series of decision trees. The computer-implemented method also includes creating a gradient boosting machine learning model as an additive function of the series of decision trees, the gradient boosting machine learning model predicting whether a further cancer patient will die within a reference number of weeks based on the additive function applied to the results from a clinical test performed on the further cancer patients and tumor types of the further cancer patients.

[0029] In a ninth aspect, an article of manufacture includes a non-transitory computer readable medium having stored thereon program instructions that, when executed by a computing device, cause the computing device to perform operations of the eighth aspect.

[0030] In a tenth aspect, a computing system includes a computer memory configured to store a gradient boosting machine learning model, the gradient boosting machine learning model trained with model training data, the model training data correlating (i) results from clinical tests performed on a first set of cancer patients and tumor types of the first set of cancer patients with (ii) whether each patient from the first set of cancer patients died within a reference number of weeks from the start of cancer treatment, and training the gradient boosting machine learning model includes (i) training a series of decision trees until a stopping condition is reached, each successive decision tree in the series of decision trees constructed based on residual values ​​of its predecessor in the series of decision trees, and (ii) determining an additive function of the series of decision trees. The computing system also includes one or more processors in communication with the computer memory and configured to execute program instructions to: (i) obtain results from laboratory tests performed on the further cancer patient; (ii) obtain a tumor type of the further cancer patient; (iii) apply a gradient boosting machine learning model to the results from the laboratory tests performed on the further cancer patient and the tumor type of the further cancer patient; and (iv) receive a prediction from the gradient boosting machine learning model of whether the further cancer patient will die within a reference number of weeks.

[0031] In an eleventh aspect, the computer-implemented method includes obtaining a gradient boosting machine learning model, the gradient boosting machine learning model being trained with model training data, the model training data including a first set of cancer patients who have undergone cancer treatment, the model training data associating (i) results from a clinical test performed on the first set of cancer patients and a tumor type of the first set of cancer patients with (ii) whether an individual patient from the first set of cancer patients died within a reference number of weeks from the start of the cancer treatment, and training the gradient boosting machine learning model includes (i) training a series of decision trees until a stopping condition is reached, each successive decision tree of the series of decision trees being constructed based on a residual value of its predecessor in the series of decision trees, and (ii) determining an additive function of the series of decision trees. The computer-implemented method also includes obtaining results from a clinical test performed on the additional cancer patients. The computer-implemented method also includes obtaining tumor types of the additional cancer patients. The computer-implemented method also includes applying the gradient boosting machine learning model to the results from the clinical test performed on the additional cancer patients and the tumor types of the additional cancer patients. The computer-implemented method also includes receiving a prediction from the gradient boosting machine learning model of whether an additional cancer patient will die within a reference number of weeks.

[0032] In a twelfth aspect, an article of manufacture includes a non-transitory computer readable medium having stored thereon program instructions that, when executed by a computing device, cause the computing device to perform operations of the eleventh aspect.

[0033] Other features and advantages of the invention will become apparent from the detailed description and claims. [Brief description of the drawings]

[0034] [Figure 1]Figure 1A shows the overall survival of patients enrolled in a phase I immuno-oncology trial (Durvalumab Study 1108 (2L+UC cohort)). The graph shows that approximately 20% of patients died early in the phase I I / O clinical trial, despite meeting the eligibility criteria. Figure 1B shows the prognostic scores published to estimate life expectancy based on retrospective studies. These scores were developed to identify patients who are likely to have a short survival time. dNLR, induced neutrophil / (leukocyte-neutrophil) ratio, LDH:lactate dehydrogenase, LLN: lower limit of normal, NLR: neutrophil / lymphocyte ratio, ULN: upper limit of normal. Arkenau, J Clin Oncol 2009; Bigot, Eur J Cancer 2017; Mezquita, JAMA Oncol 2018. [Diagram 2] Figure 2 shows the schema for the development of the high risk of early death (HREM) prognostic score. In summary, the data of all patients enrolled in three phase I trials (NCT01693562, NCT02000947, NCT02261220) were divided into two pools (training data and test data). The training data was used to build a model using 10x cross-validation. The test data was then applied to the model and the results were compared with those obtained using the RMH, GRIM and LIPI prognostic scores (see Figure 1). [Diagram 3] Figure 3A shows the proportion of patients with an overall survival (OS) score of 8 weeks or less (i.e., patients at high risk of early death). Figure 3B shows the proportion of patients with an overall survival (OS) score of 12 weeks or less (i.e., patients at high risk of early death). Sites with 7 or more patients enrolled were included (n=55 sites). By uncovering whether sites tend to recruit patients who are more or less likely to have EM, site selection can be improved by identifying patients at high risk of early death. [Figure 4] Figure 4 shows the HREM score (Figure 4A) and the associated predictive performance for identifying patients with overall survival (OS) ≦12 weeks (Figure 4B). The true positive rate (TPR) averaged about 75% in the training and testing groups. The false positive rate (FPR) was kept to about 10%. [Diagram 5] Figure 5 shows the time-dependent TPR (Figure 5A) and FPR (Figure 5B) of HREM compared with RMH, GRIM and LIPI prognostic scores. HREM better predicts patients with high risk of early death (TPR): OS ≤ 4 weeks (88.4%), OS ≤ 8 weeks (81.7%), OS ≤ 12 weeks (74.7%). HREM also suppresses FPR well. p-values: ***<0.001, **<0.01, *<0.05. [Figure 6] Figure 6 shows the OS of patients identified as at high risk of early death by HREM versus those not identified as at high risk of early death. Predicted HREM survival is significantly reduced compared to non-HREM survival. Median survival (weeks) and 95% confidence intervals (CI) are shown. [Figure 7] Figure 7 shows a comparison of OS for patients at high risk of early death as predicted by HREM with OS according to the RMH, GRIM and LIPI prognostic scores. Median survival time (weeks) and 95% confidence intervals (CI) for each score are shown. [Figure 8] Figure 8 shows the impact of PD-L1 on OS of patients predicted to be at high risk of early death. Figure 8A shows a composite graph of survival probability based on HREM, RMH, GRIM and LIPI prognostic scores for PD-L1 low / negative (solid line) and PD-L1 high (dashed line) of the patients tested. Figures 8B, 8C, 8D and 8E show the prognostic scores for PD-L1 low / negative (solid line) and PD-L1 high (dashed line) of HREM, RMH, GRIM and LIPI prognostic scores, respectively. The impact of PD-L1 was observed in patients at high risk of early death predicted by HREM, but not the other three scores. Patients at high risk of early death with low / negative PD-L1 predicted by HREM progress faster than with other scores. Median survival time (weeks) and 95% confidence interval (CI) are shown. PD-L1 status was determined in both tumor cells (TC) and immune cells (IC). *NSCLC, high: TC ≥ 25% and low / negative: TC < 25%; UC, high: TC ≥ 25% or IC ≥ 25%, low / negative: TC < 25% and IC < 25%. [Figure 9]Figure 9 shows the impact of the presence of liver metastases (mets) on the OS of patients predicted to be at high risk of early death. Figure 9A shows a composite graph of survival probability based on HREM, RMH, GRIM and LIPI prognostic scores with (solid line) and without (dashed line) liver mets for the patients tested. Figures 9B and 9C show the prognostic scores of HREM, RMH, GRIM and LIPI with (solid line) and without (dashed line) liver mets, respectively. The liver Mets positive POS population of patients predicted to be at high risk of early death from the four scores has a similar survival profile. HREM is more eligible to capture faster progressing patients in the liver Mets negative population compared to the other three scores. [Figure 10-1] FIG. 10 shows stratification of PD-L1 and HREM patients by treatment (FIGS. 10A and 10C = durvalumab; FIGS. 10B and 10D = durvalumab + tremelimumab) and cancer type (FIGS. 10A and 10B = non-small cell lung cancer (NSCLC); FIGS. 10C and 10D = bladder cancer (UBC)). [Figure 10-2] This is a continuation of Figure 10-1. [Figure 11] FIG. 11 shows that the FastProgIO prognostic score correlates with a high inflammatory proteomic signature. [Figure 12] Figure 12A shows the prognostic score of F1 model. Figure 12B highlights the comparison of HREM rate (boxed) with the same tumor type from different studies. Such comparison can help understand unexpected efficacy results. Boxed boxes show the same tumor type. [Figure 13] Figure 13 shows the variation in rapid progressor rates between study sites used in Example 4. Clearly, there was a large variation in HREM rates between study sites. Approximately 44% of study sites had HREM rates ≥ 20%. Only study sites with 7 or more patients enrolled were included (n = 79 sites). [Figure 14] FIG. 14 shows a schematic diagram for developing the F1 model. [Figure 15]Figure 15 shows the predictive performance of the F1 model using the training and test data. Similar predictive performance is seen for the training and test data. The FPR was controlled at 10% with an OS cutoff of 12 weeks. [Figure 16] FIG. 16 shows overall survival of the F1 model predicted group. [Figure 17] Figure 17 shows a comparison of OS of patients predicted to be at high risk of early death using the F1 model and the RMH, GRIM and LIPI prognostic scores. Patients predicted to be at high risk of early death by the F1 model had poorer OS compared to existing predictors. The F1 model was the only model that predicted 12-week life expectancy (i.e., the 95% CI did not overlap with the 12-week bar). [Figure 18] Figure 18 shows the impact of PD-L1 on OS for patients predicted to be at high risk of early death by the F1 model. Most patients with PD-L1 low / negative predicted to be at high risk of early death had poorer OS compared to PD-L1 high patients. Patients predicted to be at high risk of early death by the F1 model had the poorest OS compared to published scores in both PD-L1 low / negative and PD-L1 high populations. [Figure 19] Figure 19 shows the impact of liver metastasis on OS of patients predicted to be at high risk of early death by the F1 model. Patients predicted to be at high risk of early death with liver metastasis had poorer OS compared to patients without liver metastasis. F1 predicted patients at high risk of early death showed the poorest OS profile compared to published scores in both liver metastasis negative and positive populations. [Figure 20]Figure 20A shows the results of study 1108, where durvalumab alone showed promising efficacy (median OS = 4.9 months (2.6-9.1)). In the crossover study, outcomes appeared to be worse, which could be explained by F1, but study 21 showed a much worse outcome than 1108 with durvalumab (D) alone, and was also associated with poorer outcomes with durvalumab + tremelimumab (D+T) in the INFy+ cohort. Figure 20B shows that when F1 was applied to study 21, a large imbalance occurred in patients at high risk of early mortality between the cohorts. The imbalance of "poor prognosis" patients predicted by the F1 model is observed between the cohorts (i.e., not mitigated by randomization). Both the D and D+T (IFNy+) cohorts enrolled a much higher proportion of high-risk patients, providing a potential explanation for the poor outcomes in these cohorts. Other potential explanations, such as differences in baseline clinical prognostic parameters, differences in baseline biomarkers, PDL1 or IFNy+ status, did not explain the differences in outcomes. [Figure 21] Figure 21 shows the results of the ARCTIC study in patients with NSCLC, where durvalumab plus tremelimumab (D+T) showed a greater benefit in patients with PD-L1 negativity (<1%). In patients with low levels of PD-L1 (≧1%, <25%), durvalumab monotherapy and chemotherapy showed similar benefit and were better than D+T and tremelimumab alone. [Figure 22] Figure 22 shows the assessment of the imbalance between groups in F1 model patients predicted to be at high risk of early death and OS excluding patients predicted to be at high risk of early death. Similar rates of rapid progressors were predicted in all cohorts, →the difference in OS was not due to imbalances among “poor prognosis” patients. After excluding patients predicted to be at high risk of early death, D+T is still associated with better OS than chemotherapy in PD-L1 negative patients. [Figure 23] FIG. 23 shows the correlation of F1 model results with inflammatory proteomic signatures (Study 1108, UBC). [Figure 24]FIG. 24 shows the Intention-to-Treat (ITT) survival probability analysis over a 144-week period (top) and a table of at-risk patients over the same period for the three treatment schemes in the EAGLE study: chemotherapy, durvalumab and durvalumab + tremelimumab (bottom). [Diagram 25] Figure 25 shows PD-L1 high (≧25%) survival probability analysis over a 144-week period (top) and a table of at-risk patients over the same period for the three treatment schemes in the EAGLE trial: chemotherapy, durvalumab and durvalumab + tremelimumab (bottom). [Figure 26] Figure 26 shows PD-L1 low / negative (≦25%) survival probability analysis over a 144-week period (top) and a table of at-risk patients over the same period for the three treatment schemes in the EAGLE trial: chemotherapy, durvalumab and durvalumab + tremelimumab (bottom). [Figure 27] Figure 27 shows the survival probability analysis for all patients over a 144-week period (top) and a table of at-risk patients over the same period for the three treatment schemes in the EAGLE study: chemotherapy, durvalumab and durvalumab + tremelimumab (bottom). HR, hazard ratio. CI, confidence interval. Also shown are the hazard ratios (HR) and p-values ​​for the durvalumab and durvalumab + tremelimumab groups of the EAGLE study in squamous cell carcinoma of the head and neck. [Figure 28] Figure 28 shows the Intention-to-Treat (ITT) survival probability analysis over a 144-week period (top) and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Figure 29] Figure 29 shows the Intention-to-Treat (ITT) survival probability analysis excluding patients at high risk of early death over a 144-week period (top) and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 30]Figure 30 shows the survival probability analysis of PD-L1 high patients (≧25%) over a 144-week period (top) and a table of at-risk patients over the same period with the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 31] Figure 31 shows the survival probability analysis of PD-L1 high patients (>25%) excluding patients at high risk of early death over a 144-week period (top) and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 32] Figure 32 shows the survival probability analysis of PD-L1 low patients (>=1, <25%) over a 144-week period (top) and a table of at-risk patients over the same period with the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 33] Figure 33 shows the survival probability analysis (top) of PD-L1 low patients (>=1, <25%) excluding patients at high risk of early death over a 144-week period, and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 34] Figure 34 shows the survival probability analysis of PD-L1 negative patients (<1%) over a 144-week period (top) and a table of at-risk patients over the same period with the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 35] Figure 35 shows the survival probability analysis of PD-L1 negative patients (<1%) excluding patients at high risk of early death over a 144-week period (top) and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 36]Figure 36 shows the survival probability analysis of patients at high risk of early death over a 144-week period (top) and a table of at-risk patients over the same period with the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​of the F1 analysis. [Figure 37] Figure 37 shows the Intention-to-Treat (ITT) survival probability analysis excluding patients at high risk of early death over a 144-week period (top) and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Figure 38] Figure 38 shows the survival probability analysis of PD-L1 high patients (>25%) excluding patients at high risk of early death over a 144-week period (top) and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Figure 39] Figure 39 shows the survival probability analysis (top) of PD-L1 low patients (>=1, <25%) excluding patients at high risk of early death over a 144-week period, and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 40] Figure 40 shows the survival probability analysis of PD-L1 negative patients (<1%) excluding patients at high risk of early death over a 144-week period (top) and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HRs) and p-values ​​for the F1 analysis. [Diagram 41]Figure 41 shows a survival probability analysis of patients at high risk of early death determined using NLR, NEUT, ALB, LDH, GGT and AST over a 144-week period (top) and a table of at-risk patients over the same period for the three treatment schemes in the EAGLE study: chemotherapy, durvalumab and durvalumab + tremelimumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. [Diagram 42] Figure 42 shows the Intention-to-Treat (ITT) survival probability analysis (top) excluding patients at high risk of early death (as determined by NLR, NEUT, ALB, LDH, GGT and AST) over a 144-week period, and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. [Diagram 43] Figure 43 shows the survival probability analysis (top) of PD-L1 high patients (>25%) excluding patients at high risk of early death (as determined by NLR, NEUT, ALB, LDH, GGT and AST) over a 144-week period, and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. [Diagram 44] Figure 44 shows the survival probability analysis (top) of PD-L1 low patients (>=1, <25%) excluding patients at high risk of early death (as determined by NLR, NEUT, ALB, LDH, GGT and AST) over a 144-week period, and a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. [Diagram 45] Figure 45 shows the survival probability analysis (top) of PD-L1 negative patients (<1%) excluding patients at high risk of early death (as determined by NLR, NEUT, ALB, LDH, GGT and AST) over a 144-week period, as well as a table of at-risk patients over the same period for the two treatment schemes in the EAGLE study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. [Figure 46] Figure 46 shows a survival probability analysis of patients at high risk of early death determined using NLR, NEUT, ALB, LDH, GGT and AST over a 144-week period (top) and a table of at-risk patients over the same period for the three treatment schemes in the MYSTIC study: chemotherapy, durvalumab and durvalumab + tremelimumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. [Figure 47] Figure 47 shows the survival probability analysis of PD-L1 high patients (≧25%) over a 144-week period (top) and a table of at-risk patients over the same period with two treatment schemes in the MYSTIC study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. The arrows indicate the intersection of the Kaplan-Meier (KM) curves. [Figure 48] Figure 48 shows the survival probability analysis (top) of PD-L1 high patients (>25%) excluding patients at high risk of early death (as determined by NLR, NEUT, ALB, LDH, GGT and AST) over a 144-week period, and a table of at-risk patients over the same period for the two treatment schemes in the MYSTIC study: chemotherapy and durvalumab (bottom). Also shown are the hazard ratios (HR) and p-values ​​for the F1 analysis. [Figure 49] FIG. 49 shows overall survival (OS) in the treatment arms of studies 1108, HAWK, CONDOR and EAGLE. [Figure 50] FIG. 50 shows patients predicted by F1 to be at high risk of early death (FastProg) and patients not at risk of early death (non-FastProg) in the treatment arms of study 1108, HAWK, CONDOR and EAGLE. [Figure 51] Figure 51 shows Kaplan-Meier estimates for overall PD-L1 positive patients alive at 12 weeks. [Figure 52]Figure 52 shows the Immunity Instant Index (3i) score development workflow. Model requirements (1) were defined to address current challenges in predicting early mortality. Design (2) was selected to fit the model requirements in dataset, feature and method selection. Test plan (3) was prepared to evaluate model performance following globally accepted machine learning practices and to mitigate any risks associated with the open source packages used. Development (4) followed globally accepted principles of dataset separation, k-fold cross-validation during model training, hyperparameter optimization and testing on independent datasets (5). [Figure 53] Figure 53 shows the quantification of variable importance in feature selection. A gradient boosting model was fitted using a set of variables to predict 12-week life expectancy. The first 20 variables with the highest importance values ​​are plotted. The variables were clustered into high and low importance groups using K-means clustering. Abbreviations: ALB = albumin; ALP = alkaline phosphatase; AST = aspartate aminotransferase; BMI = body mass index; CL = chloride; EOSLE = eosinophil / white blood cell ratio; GGT = gamma-glutamyltransferase; GLUC = glucose; HCT = hematocrit; LDH = lactate dehydrogenase; LYMLE = lymphocyte / white blood cell ratio; MG = magnesium; NEUT = neutrophils; NEUTLE = neutrophil / white blood cell ratio; NLR = neutrophil / lymphocyte ratio; PROT = protein; TSH = thyrotropin; TUMSZINV = tumor size; URATE = uric acid; WBC = white blood cells. [Figure 54] Figure 54 shows quantification of immune immediate index (3i) score variable importance in the final 3i score model. Variables were clustered into more and less important groups using K-means clustering. Abbreviations: ALB = albumin; AST = aspartate aminotransferase; GGT = gamma glutamyl transferase; LDH = lactate dehydrogenase; NEUT = neutrophils; NLR = neutrophil / lymphocyte ratio; NSCLC = non-small cell lung cancer; SCCHN = squamous cell carcinoma of the head and neck; UBC = urothelial bladder cancer. [Figure 55] Figure 55 shows overall survival by 3i score in the training set. Total n=2213; consisting of study 1108 (all solid tumors, n=923), study 06 (non-small cell lung cancer, n=353), study 10 (urothelial bladder cancer, n=164), study 21 (gastric cancer, n=98), ATLANTIC (non-small cell lung cancer, n=371), CONDOR (squamous cell carcinoma of the head and neck, n=193) and HAWK (squamous cell carcinoma of the head and neck, n=111). [Figure 56-1] Figures 56A and 56B show overall survival by 3i score, GRIM and LIPI models in the entire training set. Total n=2213; consisting of study 1108 (all solid tumors, n=923), study 06 (non-small cell lung cancer, n=353), study 10 (urothelial bladder cancer, n=164), study 21 (gastric cancer, n=98), ATLANTIC (non-small cell lung cancer, n=371), CONDOR (squamous cell carcinoma of the head and neck, n=193) and HAWK (squamous cell carcinoma of the head and neck, n=111). Left panel: Kaplan-Meier curves for patients identified as high risk by 3i score, GRIM or LIPI. Right panel: Kaplan-Meier curves for patients identified as low risk by 3i score, GRIM or LIPI. [Figure 56-2] This is a continuation of Figure 56-1. [Figure 57-1] Figures 57A and 57B show overall survival by 3i score, GRIM and LIPI models in MYSTIC ITT patients randomized to durvalumab. N=374, non-small cell lung cancer patients. Kaplan-Meier curves for patients identified as high risk (left panel) or poor risk (right panel) by 3i, GRIM or LIPI score. [Figure 57-2] This is a continuation of Figure 57-1. [Figure 58-1] Figures 58A-58C show Kaplan-Meier estimates of overall survival in the MYSTIC ITT population. Figure 58A shows all patients, Figure 58B shows the 3i score low risk population, and Figure 58C shows the 3i score high risk population. [Figure 58-2] This is a continuation of Figure 58-1. [Figure 58-3]This is a continuation of Figure 58-2. [Figure 59-1] Figures 59A-59C show Kaplan-Meier estimates of overall survival in the MYSTIC PD-L1≧25% subgroup. Figure 59A shows all patients, Figure 59B shows the 3i score low risk population, and Figure 59C shows the 3i score high risk population. N.Missing refers to the number of patients with at least one missing laboratory value and no 3i score available. N.3i refers to the number of patients with a 3i high risk score. [Figure 59-2] This is a continuation of Figure 59-1. [Figure 59-3] This is a continuation of Figure 59-2. [Figure 60-1] Figures 60A and 60B show overall survival by 3i score, GRIM and LIPI models in EAGLE ITT patients randomized to durvalumab. N=240 patients with squamous cell carcinoma of the head and neck. Kaplan-Meier curves for patients identified as high risk (60A) or low risk (60B) by 3i, GRIM or LIPI scores. [Figure 60-2] This is a continuation of Figure 60-1. [Figure 61-1] Figures 61A-61C show Kaplan-Meier estimates of overall survival in the EAGLE ITT population. Figure 61A shows all patients, Figure 61B shows the 3i score low risk patient population, and Figure 61C shows the 3i score high risk patient population. [Figure 61-2] This is a continuation of Figure 61-1. [Figure 61-3] This is a continuation of Figure 61-2. [Figure 62-1] Figures 62A-62C show Kaplan-Meier estimates of overall survival in the EAGLE PD-L1≧25% subgroup. Figure 62A shows all patients, Figure 62B shows the 3i score low risk patient population, and Figure 62C shows the 3i score high risk patient population. [Figure 62-2] This is a continuation of Figure 62-1. [Figure 62-3] This is a continuation of Figure 62-2. [Figure 63]Figure 63 shows box plots of input variable differences between 3i score high risk and 3i score low risk patients in MYSTIC. Abbreviations: ALB = albumin; AST = aspartate aminotransferase; GGT = gamma-glutamyltransferase; LDH = lactate dehydrogenase; NEUT = neutrophils; NLR = neutrophil / lymphocyte ratio. [Figure 64] Figure 64 shows box plots of input variable differences between 3i score high risk and 3i score low risk patients in EAGLE. Abbreviations: ALB = albumin; AST = aspartate aminotransferase; GGT = gamma-glutamyltransferase; LDH = lactate dehydrogenase; NEUT = neutrophils; NLR = neutrophil / lymphocyte ratio. [Figure 65] Figure 65 shows lower expression of T effector and higher immunosuppressive gene signature in 3i high compared to 3i low NSCLC, HNSCC and UBC patients. Overall solid tumor and study 3i high patients show 1.9-fold lower expression in T effector and 1.5-fold higher expression in immunosuppressive gene signature. The number of patients (n) in each 3i high and 3i low group and Wilcox test p-value are shown (****P<0.0001). NSCLC data were obtained from MYSTIC, ATLANTIC, CP1108 and 006 studies. HNSCC data were obtained from HAWK, CONDOR and CP1108. UBC data were obtained from study 10 and CP1108. [Figure 66] Figure 66 shows that 3i high NSCLC patients have lower expression of T effectors and higher immunosuppressive gene signatures compared to 3i low NSCLC patients, regardless of PD-L1 status. The number of patients (n) in the 3i high and 3i low groups and Wilcox test p-values ​​are shown (****P<0.0001). [Figure 67]Figure 67 shows that 3i-high NSCLC patients have lower absolute T, NK and B cell counts and higher absolute MNC MDSC counts compared to 3i-low NSCLC patients. Absolute T, NK and B cell counts are at least 24% lower in 3i-high NSCLC patients compared to 3i-low NSCLC patients (data from CP1108 and 006 studies). Absolute MNC MDSC (CD14+HLA-DRlo / -) and monocyte (CD14+HLA-DR+) counts were significantly 138.2% higher in 3i-high vs. 3i-low NSCLC patients (Mystic), but no difference was observed in absolute monocyte counts. The number of patients (n) in the 3i-high and 3i-low groups and Wilcox test p-values ​​are shown (***P<0.001; ****P<0.0001). [Figure 68] Figure 68 shows that 3i high NSCLC patients have lower absolute numbers of naive T cells, activated T cells, effector memory T cells, and central memory T cells compared to 3i low NSCLC patients. Data are from CP1108 and 006 studies. The number of patients (n) in the 3i high and 3i low groups and Wilcox test p values ​​are shown (***P<0.001; ****P<0.0001). [Figure 69] Figure 69 shows that serum expression levels of CRP, IL6, and IL8 are lower in 3i-high patients compared to 3i-low NSCLC, HNSCC, and UBC patients. The number of patients (n) in the 3i-high and 3i-low groups and Wilcox test p-values ​​are shown (****P<0.0001). NSCLC data was obtained from the CP1108 study. HNSCC data was obtained from HAWK, CONDOR, and CP1108. UBC data was obtained from CP1108. [Figure 70] FIG. 70 illustrates an example of a computing device that can be configured to train and / or execute a gradient boosting machine learning model. [Figure 71] FIG. 71 illustrates an example of a computer cluster that can be configured to train and / or run a gradient boosting machine learning model. [Figure 72]Figure 72 shows a flowchart representation of an example of training a gradient boosting machine learning model. [Figure 73] Figure 73 shows a flowchart representation of an example implementation of a gradient boosting machine learning model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0035] Unless otherwise defined, all technical and scientific terms used herein have the meanings commonly understood by those skilled in the art to which the present invention belongs. The following references provide those skilled in the art with general definitions of many of the terms used in the present invention: Singleton, et al., Dictionary of Microbiology and Molecular Biology (2nd ed. 1994); The Cambridge Dictionary of Science and Technology (Walker ed., 1988); The Glossary of Genetics, 5th Ed., R. Rieger, et al. (eds.), Springer Verlag (1991); and Hale & Marham, The Harper Collins Dictionary of Biology (1991). As used herein, the following terms have the meanings ascribed to them below, unless otherwise specified.

[0036] By "anti-PD-L1 antibody" is meant an antibody that selectively binds to a PD-L1 polypeptide. Exemplary anti-PD-L1 antibodies are described, for example, in U.S. Patent Nos. 8,779,108, 9,493,565, and 10,400,039, which are incorporated herein by reference. Durvalumab (MEDI4736) or "Durva" is an exemplary anti-PD-L1 antibody suitable for the methods described herein.

[0037] "Anti-CTLA4 antibody" refers to an antibody that selectively binds to a CTLA4 polypeptide. Exemplary anti-CTLA4 antibodies are described, for example, in U.S. Patent Nos. 6,682,736, 7,109,003, 7,123,281, 7,411,057, 7,824,679, 8,143,379, 7,807,797 and 8,491,895 (wherein tremelimumab is 11.2.1), which are incorporated herein by reference. Tremelimumab or "Treme" is an exemplary anti-CTLA4 antibody.

[0038] As used herein, the terms "biomarker," "marker," or "clinicopathological marker" (which may be used interchangeably) generally refer to a protein, nucleic acid molecule, clinical indicator, or other analyte associated with a disease. In one embodiment, a clinicopathological marker is differentially expressed (or present) in a biological sample obtained from a subject with a disease (e.g., bladder cancer) relative to the concentration present in a control sample or standard control.

[0039] In another embodiment, the clinicopathological marker is an indicator of the extent of disease. For example, if the concentration of a clinicopathological marker in a patient with a disease (e.g., cancer) is higher than that of a control (e.g., a subject without cancer), the higher concentration can indicate the progression of the patient's disease. Similarly, if the concentration of a clinicopathological marker in a patient with a disease (e.g., cancer) is higher than that of an earlier measurement of the marker in the patient, the higher concentration can indicate the progression of the patient's disease. Furthermore, in some embodiments, the concentration of a clinicopathological marker can indicate the relative health (e.g., body mass index) and / or the general physiological resilience (e.g., age) of the patient.

[0040] In some embodiments, the concentration of one or more clinicopathological markers may indicate a patient's immune fitness, e.g., the relative ability of a patient's immune system to fight a particular disease by itself or to be enhanced or modified by treatment with a therapeutic agent, such as an immune checkpoint inhibitor (ICI), to fight a particular disease, such as cancer.

[0041] In this disclosure, "comprises," "including," "containing," "having," and the like may have the meaning ascribed to them in U.S. patent law, and may mean "comprise," "comprising," and the like; "consisting essentially of" or "consisting essentially of" likewise have the meaning ascribed to them in U.S. patent law, the terms being open-ended and permitting the presence of more than what is recited, but excluding prior art embodiments, so long as the basic or novel characteristics recited are not changed by the presence of more than what is recited.

[0042] As used herein, the terms "determine," "assess," "assay," "measure," and "detect," as well as "identify," refer to both quantitative and qualitative determinations, and thus the term "determine" is used synonymously herein with "assay," "measure," etc. When a quantitative determination is intended, the phrase "determine the amount" of an analyte, substance, protein, etc. is used. When a qualitative and / or quantitative determination is intended, the phrase "determine the concentration" of an analyte or "detect" an analyte is used.

[0043] "Disease" refers to any condition or disorder that damages, interferes with, or dysregulates the normal function of a cell, tissue, or organ. In diseases such as cancer (e.g., bladder cancer), the normal function of a cell tissue or organ is disrupted, allowing immune evasion and / or escape.

[0044] In the context of treatment, "responsive" means susceptible to treatment. "Subject" means a mammal, including, but not limited to, a human, such as a human patient, a non-human primate, or a non-human mammal, such as a bovine, equine, canine, ovine, or feline.

[0045] As used herein, the terms "treat," "treating," "treatment," and the like refer to reducing, alleviating, diminishing, ameliorating, arresting, or ameliorating a disorder and / or symptoms associated therewith. Although not exclusive, it will be understood that treating a disorder or condition does not require the complete elimination of the disorder, condition, or symptoms associated therewith.

[0046] Unless otherwise stated or clear from the context, the term "or" as used herein is understood to be inclusive. Unless otherwise stated or clear from the context, the terms "a," "an," and "the" as used herein are understood to be singular or plural. Similarly, when a particular term is expressed in the singular, the same term expressed in the plural is also intended, and vice versa. For example, the term "drug" also intends "drugs," and vice versa.

[0047] Unless otherwise stated or clear from the context, the term "about" as used herein is understood to mean within the normal tolerance in the art, for example, within 2 standard deviations of the mean. The term "about" is understood to mean within 5%, 10%, 9%, 8%, 7%, 6%, 5%, 4%, 3%, 2%, 1%, 0.5%, 0.1%, 0.05% or 0.01% of the stated value. Unless otherwise clear from the context, all numerical values ​​provided herein are modified by the term about.

[0048] The recitation of a listing of chemical groups in any definition of a variable herein includes definitions of that variable as any single group or combination of listed groups. The description of an embodiment of a variable or aspect herein includes that embodiment as any single embodiment or in combination with any other embodiment or portion thereof.

[0049] early death As described herein, the invention features, in one embodiment, a method of predicting whether a cancer patient is an rapid progressor or at risk of early mortality (EM; patients with an overall survival (OS) of 24 weeks or less) by taking blood samples or other samples and physiological measurements, measuring concentrations of clinicopathological markers, and creating a composite score of marker concentrations (wherein the clinicopathological markers can be, for example, neutrophils, hemoglobin, aspartate aminotransferase, and alkaline phosphatase), as described elsewhere herein.

[0050] In certain embodiments, EM is defined as patients with overall survival (OS) of 12 weeks or less.

[0051] In one embodiment, any number of clinicopathological markers can be considered, such as those commonly measured to determine the relative disease state of cancer patients and those commonly measured to determine the relative health of individuals, whether healthy or diseased.

[0052] In some embodiments, commercially available tests can be used to measure clinical pathological marker concentrations, such as antibody-based assays that use antibodies specific for the marker of interest to target such markers. Binding of the antibody to the specific marker can be detected, and the concentration and / or concentration of the marker of interest can be determined by the relative amount of antibody binding. Additional tests can be used, such as tests that measure the relative abundance of specific cell types, e.g., flow cytometers. Additional tests that can be used herein include quantitative PCR and similar techniques that provide information on relative gene expression levels. Additional tests are also contemplated herein, as recognized by those skilled in the art, that measure various clinical pathological markers for the relative health of an individual or for a particular disease state.

[0053] In some embodiments, the samples used may include tissue biopsies, cerebrospinal fluid, lymphatic fluid, whole blood, blood fractions such as serum or specific blood cell populations, urine, sweat, tears, saliva and / or feces.

[0054] In one embodiment, patients can be evaluated upon arrival at a clinical facility for clinical trial enrollment. In this case, blood can be drawn using standard techniques, and the patient's clinical pathological markers (e.g., neutrophils, hemoglobin, aspartate aminotransferase, alkaline phosphatase, etc.) can be evaluated, and the patient's overall survival can be predicted over a given period, such as 4, 8, 12 or 24 weeks, 4, 6, 8 or 10 months, 1, 2, 3, 4 or 5 years.

[0055] In another embodiment, where the centralized software platform is capable of evaluating a patient's clinical pathological markers (e.g., neutrophils, hemoglobin, aspartate aminotransferase, alkaline phosphatase, etc.) and determining overall survival, a patient can be pre-approved or remotely recruited or informed about accelerated therapy by having laboratory reports reviewed remotely.

[0056] In some embodiments, the present invention provides a method of conducting a clinical trial of an anti-cancer therapeutic, comprising the steps of identifying a candidate for the clinical trial, the candidate having cancer, determining the blood levels of i) neutrophils, ii) hemoglobin, iii) aspartate dehydrogenase, and iv) alkaline phosphatase of the candidate, calculating the predicted overall survival of the candidate, administering a first anti-cancer therapeutic to the candidate whose predicted overall survival is 12 weeks or longer and administering a second anti-cancer therapeutic to the candidate whose predicted overall survival is 12 weeks or shorter, and reducing at least one of the rate of cancer cell division, tumor growth, tumor size, tumor density, or rate of metastasis of the candidate.

[0057] In another embodiment, the invention provides a method of treating cancer in a subject in need thereof, comprising the steps of obtaining a sample from the subject, determining the concentrations of i) neutrophils, ii) hemoglobin, iii) aspartate dehydrogenase, and iv) alkaline phosphatase in the sample, calculating a predicted overall survival score for the subject, administering to the subject a first anti-cancer therapeutic if the subject's predicted overall survival is 12 weeks or greater, or administering to the subject a second anti-cancer therapeutic if the subject's predicted overall survival is less than 12 weeks, and reducing at least one of the rate of cancer cell division, tumor growth, tumor size, tumor density, or rate of metastasis in the subject.

[0058] In some embodiments, the present invention provides a method of conducting a clinical trial of an anti-cancer therapeutic, comprising the steps of identifying a candidate for the clinical trial, the candidate having cancer, determining the blood levels of i) neutrophil / lymphocyte ratio, ii) neutrophils, iii) albumin, iv) lactate dehydrogenase, v) gamma-glutamyltransferase, and vi) aspartate dehydrogenase of the candidate, calculating the predicted overall survival of the candidate, administering a first anti-cancer therapeutic to the candidate having a predicted overall survival score of 12 weeks or more, and administering a second anti-cancer therapeutic to the candidate having a predicted overall survival score of 12 weeks or less, and reducing at least one of the rate of cancer cell division, tumor growth, tumor size, tumor density, or rate of metastasis of the candidate.

[0059] In another embodiment, the present invention provides a method of treating cancer in a subject in need thereof, comprising the steps of obtaining a sample from the subject, determining the concentrations of i) neutrophil / lymphocyte ratio (NLR) protein, ii) neutrophil, iii) albumin, iv) lactate dehydrogenase, v) gamma-glutamyltransferase, and vi) aspartate dehydrogenase in the sample, calculating a predicted overall survival score for the subject, administering a first anti-cancer therapeutic to the subject if the predicted overall survival of the subject is 12 weeks or greater, or administering a second anti-cancer therapeutic to the subject if the predicted overall survival of the subject is less than 12 weeks, and reducing at least one of the rate of cancer cell division, tumor growth, tumor size, tumor density, or rate of metastasis in the subject.

[0060] In some embodiments where the patient's overall survival is predicted to be less than 12 weeks from the date of evaluation, an accelerated and / or more aggressive treatment regimen may be prescribed that is different from that intended in the particular clinical trial to which the patient was participating when evaluated for overall survival.

[0061] Immune Checkpoint Inhibitors and Early Mortality At the forefront of the ICI class are antibodies targeting programmed cell death ligand-1 (PD-L1), which show therapeutic activity by restoring antitumor T cell responses. An example of the great impact shown by ICIs is the 23% 5-year overall survival observed in patients with PD-L1-high advanced or metastatic NSCLC treated with pembrolizumab (Garon, et al., 2019). However, during the development of these agents, the EM phenomenon has been frequently observed when ICIs are administered as monotherapy for metastatic disease, occurring in the early months of treatment in clinical trials across multiple tumor types, including NSCLC (Borghaei, et al., 2015; Rizvi, et al., 2018; Socinski, et al., 2018; Mok, et al., 2019), UC (Bellmunt, et al., 2017; Powles, et al., 2018), HNSCC (Ferris, et al., 2016) and gastric cancer (Shitara, et al., 2018). Early deaths have been observed regardless of whether the specific antibody targets PD-1 or PD-L1 (Winquist, et al., 2018; Mulkey, et al., 2019), suggesting a class effect. Although this phenomenon has been widely described, there remains an urgent need for objective measures or tools to better predict patient risk for EM and aid in directing treatment to those patients most likely to have a favorable benefit:risk profile when treated with ICI therapy.

[0062] Prediction of early deaths in clinical trials investigating ICIs To ensure a comprehensive analysis across tumor types, mortality within 4, 8, and 12 weeks of first investigational treatment was first evaluated in recent Phase I-III trials. These trials included: Study 1108 (CD-ON-MEDI4736-1108), a Phase 1 study evaluating the safety, tolerability, and pharmacokinetics of MEDI4736 in patients with advanced solid tumors (n=1002) Study 06 (D4190C0006), a Phase 1b, open-label study evaluating the safety and tolerability of MEDI4736 in combination with tremelimumab in patients with advanced non-small cell lung cancer (n=379 patients) Study 10 (D4190C00010), a Phase 1b / 2, multicenter, open-label study evaluating the safety, tolerability, and efficacy of MEDI4736 in combination with tremelimumab in patients with advanced malignancies (n=168 patients) · Study 21 (D4190C00021), a phase 1b / 2 study of MEDI4736 in combination with tremelimumab, MEDI4736 monotherapy and tremelimumab monotherapy in patients with metastatic or recurrent gastric or gastroesophageal junction adenocarcinoma (n=113 patients). · CONDOR (D4193C00003), a phase II multicenter, single-arm, international study of MEDI4736 monotherapy, tremelimumab monotherapy and MEDI4736 in combination with tremelimumab in patients with recurrent or metastatic squamous cell carcinoma of the head and neck (N=263). · HAWK (D4193C00001), a phase II multicenter, single-arm, international study of MEDI4736 monotherapy in patients with recurrent or metastatic squamous cell carcinoma of the head and neck (N=112). ATLANTIC (D4191C00003), a phase II, open-label, multicenter, international study of MEDI4736 in patients with locally advanced or metastatic non-small cell lung cancer (stage IIIB-IV) previously treated with at least two prior systemic therapies, including one platinum-based chemotherapy regimen (N=371). MYSTIC (D419AC00001), a Phase III, randomized, open-label, multicenter, international study (N=1118) comparing MEDI4736 in combination with tremelimumab or MEDI4736 monotherapy with platinum-based standard of care chemotherapy in the first-line treatment of patients with advanced or metastatic non-small cell lung cancer.

[0063] EAGLE (D4193C00002), a Phase III, randomized, open-label, multicenter, international study (N=736) comparing MEDI4736 monotherapy and MEDI4736 in combination with tremelimumab to standard of care in patients with recurrent or metastatic squamous cell carcinoma of the head and neck.

[0064] Combining these studies, a range of 17.84% to 30.36% of patients treated with ICIs were observed to die ≦12 weeks after the first day of study treatment (Table 9).

[0065] JPEG2025041587000001.jpg83170

[0066] Overall, the likely higher incidence of EM in patients treated with ICIs compared with active comparators indicates the need for improved prognostic methodology to better identify patients who are unlikely to survive beyond 12 weeks and therefore unlikely to benefit from ICI treatment.

[0067] Existing prognostic scores predict early death To improve the accuracy of physicians' prognostic predictions for patients with advanced or metastatic cancer, many studies have been conducted (Ploquin, et al., Crit Rev Oncol Hematol. 2012), contributing to the development of prognostic scores for use in clinical facilities. Most of these studies identified specific clinical and laboratory findings associated with poor outcomes in clinical trials. Several issues have reduced the usefulness of these studies in addressing the phenomenon of early death. First, some studies relied on subjective and investigator-dependent parameters such as performance status (i.e., Eastern Cooperative Oncology Group ("ECOG"), Karnofsky index) (Krishnan M; J Support Oncol 2013). Another limitation of these clinical scores is the lack of validation in large independent studies. Finally, few of these scores were developed in the ICI era, and in this context, their relevance remains unclear. An overview of these scores is provided below.

[0068] Royal Marsden Hospital (RMH) Score The RMH score was one of the first scores to predict early life expectancy in oncology. It was initially developed in the context of phase I trials investigating cytotoxic and targeted therapies. The RMH score is based on the following variables: lactate dehydrogenase (LDH, > upper limit of normal [ULN]), albumin (< 35 g / L) and number of metastatic sites (> 2) (Arkenau, et al., 2009). Each variable is awarded 1 point. If the sum of these variables exceeds 1, the patient is classified as having a poor prognosis according to RMH. Application of the RMH score in a large retrospective cohort (N > 1000 patients) of patients, regardless of tumor type (Wheler, et al., 2012) confirmed the poor prognosis associated with an RMH score > 1. High-risk RMH was associated with a poor prognosis in ICI-related retrospective studies of patients (Bigot, et al., 2017; Minami et al., 2019).

[0069] Gustave Roussy Immunology (GRIm) Score The GRIm-score is a more recent attempt to improve patient life expectancy estimation in the context of oncology phase I trials investigating ICIs. It is based on albumin (<35g / l), LDH (>ULN) and neutrophil / lymphocyte ratio (NLR;>6) (Bigot, et al. 2017). As with the RMH score, each variable is assigned 1 point. ICI-treated patients with a sum of variables >1 have been associated with poor overall survival in retrospective patient pools (Bigot, et al. 2017; Minami, et al. 2019).

[0070] Lung Immune Prognostic Index (LIPI) score The LIPI score was developed to predict the outcome of NSCLC patients treated with ICIs in various settings, including routine clinical care, expanded access, compassionate use programs, and clinical trials. It is based on the following variables: induced neutrophil / (leukocyte-neutrophil) ratio (dNLR) >3 and LDH >ULN. Each variable is awarded 1 point. NSCLC patients with a sum of variables >1 are associated with a poor prognosis in LIPI. In the first publication, the LIPI score was applied retrospectively in two independent cohorts of NSCLC patients treated with ICIs (N=305 patients) or chemotherapy (N=162 patients) (Mezquita et al., 2018). The authors found that a LIPI score >1 was associated with poor overall survival in ICI-treated patients, but not in chemotherapy-treated patients. However, a subsequent analysis of LIPI conducted by the FDA on pooled clinical testing data from a study evaluating 1368 patients with second-line metastatic NSCLC who received ICIs and 1072 patients who received chemotherapy did not confirm these results (Kazandjian, et al. 2018). This study showed that LIPI may affect prognosis regardless of treatment (ICI or chemotherapy) for second-line metastatic NSCLC. Finally, another study investigated the LIPI score in relation to metastatic patients with various solid tumors enrolled in a phase 1 trial. Although this analysis was performed on a very heterogeneous cohort, a LIPI score >1 was also associated with poor overall survival, suggesting that its prognostic role is not limited to NSCLC patients (Varga, et al. 2019).

[0071] Immune fitness All six 3i blood test variables (NLR, neutrophils, albumin, LDH, GGT, AST) assessed by standard laboratory criteria and identified by feature selection have been previously reported to be prognostic in various tumor types. For example, high NLR has been associated with poor prognosis in many tumor types, including gastrointestinal (Bowen, et al. 2017), melanoma (Ding, et al. 2018), NSCLC (Fukui, et al. 2019), UC (Sacdalan, et al. 2018) and HNSCC (Yu, et al. 2018, Tham, et al. 2018). This ratio likely reflects an assessment criterion of immune fitness that may affect tumor immune responses, and the association with poor prognosis is determined to be independent of the type of ICI. Fukui et al. observed a significant association between high NLR and poor prognosis in NSCLC patients treated with nivolumab. In a meta-analysis of ICI studies in melanoma, NSCLC and UC, Sacdalan et al. observed an association between high NLR and poor outcome, suggesting its potential as a prognostic marker. It has also been speculated that tumor burden and chronic inflammation may contribute to high NLR and subsequently poor prognosis (Bigot, et al. 2017). Absolute neutrophil count (ANC) is a surrogate marker of physiological stress, including inflammation, and in NSCLC patients, high ANC after nivolumab treatment has been observed to be independently associated with poor overall survival (Khunger, et al. 2018). Albumin (ALB) is a well-known marker of malnutrition and cachexia in cancer patients, and low ALB has been observed to be associated with poor overall survival in many malignancies, including SCCHN, breast cancer, lung cancer and gastrointestinal cancer (Gupta, et al. 2010). ALB production is also decreased in association with chronic inflammation.

[0072] LDH is an established marker reflecting enhanced cell turnover primarily in malignant disease, but also in non-malignant diseases (e.g., hemolytic anemia). A recent FDA analysis of early deaths in ICI studies reported elevated LDH as a significant risk factor primarily in melanoma patients (Mulkey, et al. 2019). In a systematic review of 76 studies, high LDH was associated with poor prognosis in renal, melanoma, gastric, prostate, nasopharyngeal and lung cancers (Petrelli, et al. 2015). In advanced cancers, elevated circulating LDH concentrations reflect increased metabolic activity and rapid cell proliferation that results in increased leakage of the intracellular enzyme LDH into the blood circulation.

[0073] In cancer patients, abnormalities in liver function tests such as AST and GGT should be closely monitored. Abnormalities in liver function tests not only reflect existing disease, but also cancer progression / liver metastasis and drug toxicity, all of which are known causes of death in cancer patients. Therefore, AST and GGT have been reported as prognostic markers in cancer (Freis, et al. 2017; Luo, et al. 2017), and GGT has been suggested to play a role in tumor initiation, invasion and drug resistance.

[0074] How to Identify Immunocompromised Patients The 3i score can be used to identify patients who may have a high probability of EM and therefore lack immune fitness. Lack of immune fitness suggests that such patients should postpone ICI treatment until their immune fitness can be improved by effective treatment. Thus, therapeutic interventions that enhance immune system fitness in 3i high-risk patients may enable these patients to receive ICI treatment with reduced EM risk, further increasing the chances of therapeutic benefit of ICI treatment.

[0075] Similarly, methods for predicting the efficacy of ICI treatment are contemplated herein. For example, patients determined to be 3i high-risk patients are predicted not to be effectively treated with ICIs such as durvalumab.

[0076] Type of Cancer The types of cancer contemplated for treatment herein include, for example, NSCLC, advanced solid malignancies, bile duct tumors, bladder cancer, colorectal cancer, diffuse large B-cell lymphoma, esophageal tumors, esophageal squamous cell carcinoma, extensive stage small cell lung cancer, gastric adenocarcinoma, gastric cancer, gastroesophageal junction cancer, head and neck cancer, head and neck squamous cell carcinoma, hepatocellular carcinoma, Hodgkin's lymphoma, lung cancer, melanoma, mesothelioma, metastatic clear cell renal carcinoma, metastatic melanoma, metastatic noncutaneous melanoma, multiple myeloma, nasopharyngeal carcinoma, non-Hodgkin's lymphoma, ovarian cancer, fallopian tube cancer, peritoneal tumors, pleural mesothelioma, prostate tumors, recurrent or metastatic PD-L1 positive or negative SCCHN, recurrent squamous cell lung cancer, renal cell cancer, renal cell carcinoma ... carcinoma, SCCHN, pharynx pituitary squamous cell carcinoma, laryngeal squamous cell carcinoma, small cell lung cancer, squamous cell carcinoma of the head and neck, squamous cell lung cancer, TNBC, transitional cell carcinoma, unresectable or metastatic melanoma, urothelial cancer and urothelial carcinoma. In certain embodiments, contemplated treatment regimens may include a biologic component, such as an antibody, and a chemotherapeutic component. Contemplated antibodies include anti-PD-L1 antibodies, such as durvalumab (MEDI4736), nivolumab, pembrolizumab, avelumab, atezolizumab, KNO35, anti-PD-1 antibodies, such as REGN2810, SHR1210, IBI308, PDR001, anti-PD-1, BGB-A317, BCD-100 and JS001, and anti-CTLA4 antibodies, such as tremelimumab or ipilimumab. Additional antibodies are also contemplated herein. Therapeutically effective antibody subparts are also contemplated herein.

[0077] Treatment method In one embodiment, the present disclosure is directed to effectively treating cancer patients with novel anti-cancer therapeutics and providing cancer patients with greater utilization of existing therapies. For example, the present disclosure contemplates treating cancer patients with immune checkpoint inhibitors (ICIs), such as those targeting the PD-1 / PD-L1 axis (PDX), and other IO therapies, such as immune system agonists.

[0078] Anticancer Therapeutic Agents Information regarding durvalumab (or a fragment thereof) for use in the methods provided herein can be found in U.S. Patent Nos. 8,779,108, 9,493,565 and 10,400,039, the disclosures of which are incorporated herein by reference in their entireties. In certain embodiments, durvalumab or an antigen-binding fragment thereof for use in the methods provided herein comprises the CDR sequences of the variable heavy and variable light chains of the 2.14H9OPT antibody disclosed in the above U.S. patents.

[0079] Information regarding tremelimumab (or an antigen-binding fragment thereof) for use in the methods provided herein can be found in U.S. Pat. No. 6,682,736, where tremelimumab is referred to as 11.2.1, the disclosure of which is incorporated herein by reference in its entirety.

[0080] The amount of anti-cancer therapeutic agent, such as an antibody or antigen-binding fragment thereof, to be administered to a patient will depend on a variety of parameters, such as the patient's age, weight, clinical evaluation, tumor burden, and / or other factors, including the judgment of the attending physician. Acceptable routes of administration are contemplated, including, but not limited to, intravenous (e.g., intravenous infusion), parenteral, or subcutaneous routes of administration.

[0081] Additional therapeutic agents (chemotherapeutic or biologic) contemplated herein include cisplatin / gemcitabine or methotrexate, vinblastine, ADRIAMYCIN, TM(doxorubicin), cisplatin (MVAC), carboplatin-based regimens or single agent taxanes or gemcitabine, temozolomide or dacarbazine, vinflunine, docetaxel, paclitaxel, nab-paclitaxel, vemurafenib, erlotinib, afatinib, cetuximab, bevacizumab, erlotinib, gefitinib and / or pemetrexed. Further examples include drugs that target the DNA damage repair system, such as poly(ADP-ribose) polymerase 1 (PARP1) inhibitors and therapeutic agents that inhibit WEE1 protein kinase activity, ATR protein kinase activity, ATM protein kinase activity, Aurora B protein kinase activity and DNA-PK activity.

[0082] The therapeutic compositions or methods contemplated herein may be combined with any one or more of the other therapeutic compositions and methods provided herein.

[0083] Companion Diagnostics In one embodiment, the present disclosure contemplates a device comprising one or more laboratory tests as described elsewhere herein for measuring concentrations of clinical pathological markers in a patient sample applied to the device. The device may further comprise a computing system comprising a gradient boosting machine learning model that predicts whether a patient has a likelihood of dying within a given time period based on the measured concentrations of the clinical pathological markers from the one or more laboratory tests.

[0084] In another embodiment, a system is contemplated that includes one or more testing devices directly or indirectly linked to one or more computing devices, where once the concentration of one or more clinical pathological markers is determined, the data is transmitted to the computing device and provided to a gradient boosting machine learning model within the computing device, which can then output predictions regarding patient EM and / or immune fitness or predictions of efficacy of an ICI therapeutic, such as durvalumab, to another component of the system.

[0085] In some embodiments, clinical pathological marker concentrations can be measured using commercially available tests, such as antibody-based assays that target a marker of interest using an antibody specific for such a marker. Binding of the antibody to the specific marker can be detected, and the concentration and / or concentration of the marker of interest can be determined by the relative amount of antibody binding.

[0086] In some embodiments, the samples used may include tissue biopsies, cerebrospinal fluid, lymphatic fluid, whole blood, blood fractions such as serum or specific blood cell populations, urine, sweat, tears, saliva and / or feces.

[0087] In one embodiment, patients can be evaluated upon arrival at a clinical facility for clinical trial enrollment. In this case, blood can be drawn using standard techniques, and the patient's clinical pathological markers (e.g., neutrophils, hemoglobin, aspartate aminotransferase, alkaline phosphatase, etc.) can be evaluated, and the patient's overall survival can be predicted over a given period, such as 4, 8, 12 or 24 weeks, 4, 6, 8 or 10 months, 1, 2, 3, 4 or 5 years.

[0088] In another embodiment, where the centralized software platform is capable of evaluating a patient's clinical pathological markers (e.g., neutrophils, hemoglobin, aspartate aminotransferase, alkaline phosphatase, etc.) and determining a composite score, a patient can be pre-approved or remotely recruited or informed about accelerated therapy by having the lab reports reviewed remotely. EXAMPLES

[0089] The following examples are put forth so as to provide those of ordinary skill in the art with a complete disclosure and description of how to practice the invention, and are not intended to limit the scope of what the inventors regard as their invention.

[0090] Example 1 – Validation of a prognostic score (High Risk of Early Death; HREM) to improve patient selection for immuno-oncology trials This example provides an overview of a new prognostic scoring method for use in cancer immunology clinical trials. As used herein, "High Risk of Early Death", "HREM" and "FastProgIO" are interchangeable.

[0091] background Estimation of life expectancy is crucial for patient selection in cancer clinical trials, especially in phase I trials. However, most trial protocols use subjective eligibility criteria to achieve this goal (e.g., life expectancy of 12 weeks). Although eligibility criteria help determine which patients may be included, they do not provide a reliable criterion for determining overall survival (see Figure 1A). Although several prognostic scores have been developed from retrospective studies to help predict patients' EM and OS (Figure 1B), their utility in the immuno-oncology (IO) context remains unknown. The observation of a high proportion of patients with EM (overall survival [OS] ≤ 12 weeks) in IO trials further highlights the need for more objective predictive criteria.

[0092] the purpose The aim of this study was to develop a new prognostic score using basic clinical tests at baseline to identify patients at high risk of early death (OS ≤ 12 weeks) using data from three phase I IO trials involving patients with non-small cell lung cancer (NSCLC) and urothelial carcinoma (UC). In addition, this study also sought to compare the proportion of patients at high risk of early death between clinical centers in these studies. We evaluated the new prognostic score (HREM) against three published prognostic scores (RMH, GRIM, and LIPI). Finally, we evaluated the impact of tumor / immune cell programmed death ligand-1 (PD-L1) expression and liver metastasis on predicting patients at high risk of early death.

[0093] method We analyzed patients with NSCLC and UC (N=972) prospectively enrolled in three phase 1 trials (NCT01693562, NCT02000947, and NCT02261220) investigating durvalumab (anti-PD-L1) ± tremelimumab (anti-cytotoxic T-lymphocyte antigen [CTLA]-4) across 160 centers. We also assessed variation in 12-week survival rates across centers.

[0094] A composite score of clinicopathological variables (HREM) predicting 12-week survival was developed in 648 patients and validated in 324 patients using multiple linear regression (Figure 2). Blood samples collected from patients were used to measure neutrophil, hemoglobin, aspartate aminotransferase, and alkaline phosphate concentrations using commercially available assays.

[0095] HREM performance (prediction of patients at high risk of early death with overall survival ≤12 weeks) was compared with that of existing published scores (see Figure 1) in terms of time-dependent true positive rate (TPR) and false positive rate (FPR).

[0096] result The rate of rapid progressors among patients analyzed per study is shown in Table 1 (see also Figures 3A and 3B).

[0097] JPEG2025041587000002.jpg87170

[0098] The baseline characteristics of NSCLC and UC patients (shown in Table 2 ) were similar.

[0099] JPEG2025041587000003.jpg204170

[0100] HREM score The HREM score included neutrophil count, aspartate aminotransferase, alkaline phosphatase, and hemoglobin as predictive markers. The true positive rate (TPR) for identifying patients with OS ≤ 12 weeks was 74.7% (74.1% in the training dataset and 75.9% in the test dataset), while the false positive rate (FPR) was approximately 10% (Figure 4). The time-dependent TPR of HREM was superior to or similar to that of RMH, GRIM, and LIPI across the OS range of 4 to 12 weeks (Figure 5). The time-dependent FPR of HREM was superior to GRIM, slightly lower than RMH, and higher than LIPI (Figure 5B). Patients identified as at high risk of early death by HREM and those not identified as at high risk of early death had clearly different survival profiles (Figure 6). Patients identified as at high risk of early death by HREM had shorter OS than those identified by the three published scores (Figure 7).

[0101] The impact of PD-L1 expression and liver metastases in patients at high predicted risk of early death The impact of PD-L1 was observed in patients with a high risk of early death predicted by HREM, but not in patients predicted by other scores: HREM predicted that patients with a high risk of early death with PD-L1 low / negative expression would progress faster than patients with PD-L1 high expression (Figure 8).

[0102] In patients with a high risk of early death predicted by all four scores, an effect of liver metastases was observed: patients with liver metastases (METS) progressed faster than patients without liver metastases (Figure 9). In patients without liver METS, patients with a high risk of early death predicted by HREM had poorer survival than patients predicted by the other three scores.

[0103] conclusion The proportion of patients at high risk of early death varied considerably between centers, suggesting that the use of subjective criteria to estimate 12-week life expectancy at study entry is clearly suboptimal. In this large cohort of NSCLC and UC patients, HREM outperformed the RMH, GRIM, and LIPI prognostic scores significantly in predicting patients at high risk of early death. An effect of PD-L1 was observed in patients at high risk of early death, as predicted by HREM, but not by the three published scores, indicating a more sensitive approach to patient assessment.

[0104] In patients predicted to be at high risk of early death by all four scores, an effect of liver metastases was observed. However, in patients without liver metastases, patients predicted to be at high risk of early death by HREM had shorter OS than those predicted by the published scores. The success of HREM prognostic assessment for patients with NSCLC and UC suggests that this approach may be useful in additional cancer types.

[0105] Example 2 – Evaluation of HREM in different tumor types This example demonstrates the applicability of the HREM prognostic scoring method for use in scoring other cancer types.

[0106] method The proportion of patients predicted to be at high risk of early death was analyzed among patients with various types of cancer at a Phase I trial (Study 1108) center.

[0107] result The rapid progressor rates of the analyzed patients are shown in Table 3 (see also Figure 10).

[0108] JPEG2025041587000004.jpg116170

[0109] Table 4 shows a comparison of HREM with other published scores.

[0110] JPEG2025041587000005.jpg103170

[0111] conclusion HREM can be used to predict patients at high risk of early death in multiple cancer types. Therefore, this prognostic score can be used to improve patient intervention and phase I outcomes in multiple cancer types. The strength of prediction can be increased by increasing the sample size.

[0112] Example 3 – Correlation of HREM with inflammatory proteins In this example, we investigated the potential correlation between HREM results and inflammatory proteins.

[0113] method Using the UC cohort, patient serum samples were examined for concentrations of inflammatory proteins that correlated with HREM scores by Spearman correlation.

[0114] result The HREM scores were highly correlated with 10 different inflammatory proteins (see FIG. 11).

[0115] conclusion HREM can be useful in generating treatment-related biological inferences that help inform patient care and / or Phase I trial selection.

[0116] Example 4 - Development of a pan-tumor prognostic model (F1 model) In this example, the aim was to develop an improved prognostic model (F1 model) based on the success of the HREM model.

[0117] method To develop a new prognostic model with improved true and false positive rates, patients from a combination of four studies across several cancer types were used (Table 5) (see also Figures 12 and 13).

[0118] JPEG2025041587000006.jpg179170

[0119] The development of the F1 model is shown in Figure 14. Several standard laboratory tests were included in the model, including ALB, LAP, AST, LDH, NEUT and WBC. Prediction modeling methods such as ensemble tree-based model and random forest model were used to develop the model.

[0120] result The predictive performance of the F1 model on the training and testing data is shown in Figure 15. The overall survival of the F1 model is shown in Figure 16. A comparison of the F1 model with published prognostic scores is shown in Figure 17. The impact of PD-L1 liver metastasis is shown in Figures 18 and 19, respectively. Taken together, these results indicate that the F1 model is a much better predictor of OS than previously available prognostic scores such as RMH, GRIM, and LIPI.

[0121] Example 5 - Retrospective application of the F1 model to Study 21 (gastroesophageal cancer) and ARCTIC (NSCLC) To demonstrate the utility of retrospective analysis of previous studies, the F1 model was used to review two phase I studies for gastroesophageal cancer using durvalumab alone (Study 1108) or durvalumab (D) alone, tremelimumab (T) alone, durvalumab + tremelimumab (D+T) and durvalumab + tremelimumab in the IFNy+ cohort (Study 21). The results can be seen by comparing Figures 20A and 20B.

[0122] The results of the ARCTIC test (Figure 21) were also taken into consideration. The results of applying the F1 model to the ARCTIC test are shown in Figure 22.

[0123] conclusion The F1 model is a stronger predictor of OS than the HREM model. It is superior to published scores (poor, OS, generalizability to all tumor types, impact of PD-L1). The F1 model can be used to evaluate ongoing or previous clinical trials to evaluate the imbalance in rapid progressor rates between arms that may potentially affect outcomes. The F1 model could also be incorporated into the inclusion / exclusion criteria of future phase I trials (titration phase, randomized expansion) as a more objective way to exclude high-risk patients. In addition, the F1 model could be used to improve site selection by excluding centers with a history of enrolling patients at high risk of early death. Excluding patients at high risk of early death in phase I trials (titration and randomized expansion) would likely increase the power of phase I trials to detect efficacy signals.

[0124] Example 6 – Correlation of F1 model with inflammatory proteins In this example, potential correlations between the results of the F1 model and inflammatory proteins were investigated.

[0125] method Using study 1108, the UC cohort, patient serum samples were examined for concentrations of inflammatory proteins (C-reactive protein, IL-6, TIMP-1, and IL-8) that correlate with F1 model score.

[0126] result HREM scores were significantly correlated with 10 different inflammatory proteins (see FIG. 23). Other correlated biomarkers (r>0.5) include ferritin (FRTN), tenascin-C (TN-C), 70 kilodalton heat shock protein (HSP70), necrosis factor receptor 1 (TNFR1), interleukin-2 receptor alpha (IL-2Rα), tumor necrosis factor receptor 2 (TNFR2), insulin-like growth factor binding protein 2 (IGFBP2), and myeloid progenitor inhibitory factor 1 (MPIF1).

[0127] conclusion The F1 model will be useful in generating treatment-related biological inferences that will help inform patient care and / or Phase I trial selection.

[0128] Example 7 - Training and testing of a new F1 model using the EAGLE study and the clinical pathological markers neutrophils (NEUT), lymphocyte percentage (LYM%), albumin (ALB), lactate dehydrogenase (LDH) and URATE This example aimed to test an improved prognostic model (F1 model) based on the success of the initial F1 model.

[0129] method To develop a new prognostic model with improved true and false positive rates, patients from a combination of three studies across several cancer types were used (Table 6).

[0130] JPEG2025041587000007.jpg59170

[0131] The model was developed as shown in FIG.

[0132] To test the model, the EAGLE trial was used (736 participants with head and neck squamous cell carcinoma, receiving three treatment schemes: chemotherapy, Durva alone and Durva+Treme). Overall survival for the intention-to-treat, PD-L1 high and PD-L1 low schemes of the trial are shown in Figures 24-26. Hazard ratios (HRs) are calculated from proportional Cox regressions including PD-L1 (≥25% vs <25%), tumor location / HPV status and smoking status as strata. p-values ​​are calculated from stratified log-rank tests stratified by PD-L1 (≥25% vs <25%), tumor location / HPV status and smoking status.

[0133] result The new F1 model predicted patients with high risk of early death (Figure 27). Overall survival of patients with chemotherapy and Durva scheme tested with the new F1 model is shown in Figure 28. Overall survival of patients with chemotherapy and Durva scheme excluding high risk prediction of early death is shown in Figure 29. Overall survival of all patients with PD-L1 high is shown in Figure 30, and data excluding patients predicted to be at high risk of early death is shown in Figure 31. Overall survival of all patients with PD-L1 high is shown in Figure 32, and data excluding patients predicted to be at high risk of early death is shown in Figure 33. Overall survival of all patients with PD-L1 negative is shown in Figure 34, and data excluding patients predicted to be at high risk of early death is shown in Figure 35.

[0134] Example 8 - Training and testing of new F1 models using the EAGLE study and neutrophils (NEUT), lymphocyte percentage (LYM%), albumin (ALB) and lactate dehydrogenase (LDH) as clinicopathological markers This example aimed to improve the F1 model by eliminating URATE from the clinicopathological markers used.

[0135] method To develop a new prognostic model with improved true and false positive rates, patients from a combination of three studies across several cancer types were used (Table 7). Hazard ratios (HRs) are calculated from proportional Cox regression including PD-L1 (>=25% vs <25%), tumor location / HPV status, and smoking status as strata. p-values ​​are calculated from stratified log-rank tests stratified by PD-L1 (>=25% vs <25%), tumor location / HPV status, and smoking status.

[0136] JPEG2025041587000008.jpg59170

[0137] The model was developed as shown in FIG.

[0138] To test the model, the EAGLE trial was used (736 participants with head and neck squamous cell carcinoma undergoing three treatment schemes: chemotherapy, Durva alone and Durva+Treme).

[0139] result The F1 model using neutrophils (NEUT), lymphocyte percentage (LYM%), albumin (ALB) and lactate dehydrogenase (LDH) predicted patients at high risk of early death (Figure 36). The overall survival of patients with chemotherapy and Durva scheme tested with the new F1 model, excluding patients identified as at high risk of early death, is shown in Figure 37. The overall survival of all PD-L1 high patients excluding patients predicted at high risk of early death is shown in Figure 38, the overall survival of PD-L1 low patients not including patients predicted at high risk of early death is shown in Figure 39, and the overall survival of PD-L1 negative patients not including patients predicted at high risk of early death is shown in Figure 40.

[0140] Example 9 - Training and testing of new F1 models for both non-small cell lung cancer and squamous cell head and neck cancer using the EAGLE and MYSTIC studies and the neutrophil / lymphocyte ratio (NLR), neutrophils, albumin, lactate dehydrogenase; gamma-glutamyltransferase and aspartate aminotransferase as clinical pathological markers. This example aimed to utilize knowledge gained from predictors of non-small cell lung cancer and squamous cell head and neck cancer in patients at high risk of early death to develop a single model for predicting both cancers. The non-small cell lung cancer markers evaluated are neutrophil / lymphocyte ratio (NLR); albumin (ALB), lactate dehydrogenase (LDH), gamma-glutamyltransferase (GGT) and aspartate aminotransferase (AST). The predictors of squamous cell head and neck cancer in patients at high risk of early death evaluated are neutrophil (NEUT), lymphocyte percentage (LYM%), albumin (ALB), lactate dehydrogenase (LDH) and URATE. The following six markers were selected: neutrophil / lymphocyte ratio (NLR); neutrophil (NEUT), albumin (ALB), lactate dehydrogenase (LDH), gamma-glutamyltransferase (GGT) and aspartate aminotransferase (AST).

[0141] method To develop a new prognostic model with improved true and false positive rates, patients from a combination of seven studies across several cancer types were used (Table 8). Hazard ratios (HRs) are calculated from proportional Cox regressions including PD-L1 (>25% vs <25%), tumor location / HPV status, and smoking status as strata. p-values ​​are calculated from stratified log-rank tests stratified by PD-L1 (>25% vs <25%), tumor location / HPV status, and smoking status.

[0142] JPEG2025041587000009.jpg194170

[0143] The model was developed as shown in FIG.

[0144] To test this model, the ARCTIC, EAGLE (736 participants with squamous cell carcinoma of the head and neck receiving three treatment schemes: chemotherapy, Durva alone and Durva+Treme) and MYSTIC trials (1118 participants with PD-L1-high non-small cell lung cancer receiving multiple treatment schemes: chemotherapy, Durva alone and Durva+Treme) were used.

[0145] result The F1 model using neutrophil / lymphocyte ratio (NLR), neutrophils, albumin, lactate dehydrogenase, gamma-glutamyltransferase and aspartate aminotransferase predicted patients at high risk of early death in the EAGLE study (Figure 41). The overall survival of chemotherapy and Durva scheme patients in the EAGLE study tested with this F1 model, excluding patients identified as at high risk of early death, is shown in Figure 42. The overall survival of all PD-L1 high patients in the EAGLE study excluding patients predicted to be at high risk of early death is shown in Figure 43, the overall survival of PD-L1 low patients not including patients predicted to be at high risk of early death is shown in Figure 44, and the overall survival of PD-L1 negative patients not including patients predicted to be at high risk of early death is shown in Figure 45. The F1 model predicted patients at high risk of early death in the MYSTIC study (Figure 46). Overall survival of all PD-L1 high patients in MYSTIC is shown in Figure 47, and overall survival of PD-L1 high patients in MYSTIC excluding patients predicted to be at high risk of early death is shown in Figure 48. The proportion of observed overall survival of 12 weeks or more or less than 12 weeks for several cohorts is shown in Figure 49, and patients predicted by F1 to be at high risk of early death and patients not at high risk of early death in these cohorts are shown in Figure 50.

[0146] conclusion The F1 model, using neutrophil / lymphocyte ratio (NLR), neutrophils, albumin, lactate dehydrogenase, gamma-glutamyltransferase, and aspartate aminotransferase as clinicopathological markers, is a strong predictor of OS. The F1 model can be used to evaluate ongoing or previously performed clinical trials to evaluate any imbalance in rapid progressor rates between arms that may potentially affect outcomes. In addition, the F1 model could be incorporated into the inclusion / exclusion criteria of future phase I trials (titration phase, randomized expansion) as a more objective way to exclude high-risk patients. In addition, the F1 model could be used to improve site selection by excluding centers with a history of enrolling patients at high risk of early death. Excluding patients at high risk of early death in phase I trials (titration and randomized expansion parts) would likely increase the power of phase I trials to detect efficacy signals.

[0147] Example 10 – Development of an Immediate Immunity Index (3i score) to identify patients at risk of early mortality overview A new approach: Identifying patients at high risk of early mortality in relation to ICIs Although RMH, GRIM and LIPI scores showed prognostic impact, the recent availability of a large number of patients treated with ICIs in the AstraZeneca prospective study provides an opportunity to improve the identification of patients at high risk of early death associated with ICIs compared with chemotherapy active controls.

[0148] In randomized trials comparing ICIs with active controls in advanced disease, Kaplan-Meier (KM) curve crossings for OS have been predominantly observed, with ICI arms generally showing higher mortality within the first 12 weeks after randomization. This has been shown in the MYSTIC trial comparing durvalumab with chemotherapy in first-line NSCLC. Figure 47 shows the OS curves (curve crossings indicated by arrows) for the primary analysis population of PD-L1-positive patients (defined as patients with cancer expressing PD-L1 in ≥25% of cells). Mortality during the first 12 weeks was 17.2% in the experimental arm compared with 8.0% in the control arm. In the 12- to 24-week period, these rates were 12.7% and 15.0%, respectively. Over time, mortality continued to improve more favorably for durvalumab. Figure 51 shows the KM curves for OS after excluding patients who died within the first 12 weeks. In this subset of patients who survived 12 weeks or more, crossover of the OS curves is no longer observed. Similar observations have been made in other studies comparing ICIs with chemotherapy (Borghaei, et al. 2015; Mok, et al. 2019). Therefore, based on empirical observations in these studies, to predict the risk of EM with ICIs, we defined EM as any death occurring within the first 12 weeks after randomization (in randomized trials) or after the start of study treatment (in single-arm trials).

[0149] Development of the 3i Score Given the data collected through trials examining ICIs, the challenges in objectively predicting EM, and the limitations of previously established scores, we set out to develop a prognostic model (hereafter referred to as the Immuno-Immediate Index [3i score]) to identify patients at risk for EM using machine learning. The development workflow, including model requirement definition, design, development, and testing, is described below.

[0150] Model Requirements As shown in Figure 52, a key requirement of the 3i Score is to predict EM (death within 12 weeks of randomization or treatment initiation) using machine learning, ideally utilizing continuous input values ​​obtained from some or all of the baseline laboratory tests performed routinely. Machine learning models are increasingly being used in healthcare and clinical facilities to identify subtle but predictive patterns in raw data that may be lost by discretizing the test values. Deploying machine learning models in the real world is complicated due to the need for robust performance and interfacing software. As a result, acceptable performance metrics across training, tuning, and testing as well as flexible deployment options influenced the final 3i Score.

[0151] Design choices in identifying datasets for development, separating datasets for training, tuning and testing, and feature and method selection were driven by model requirements and cross-functional expertise in machine learning and clinical research.

[0152] Developing machine learning models through training, tuning and testing The primary activities in developing a supervised machine learning model are most commonly subdivided into three key steps: training, tuning, and testing.

[0153] During training, the model's parameters, such as tree depth, are optimized to encode the relationship between inputs and outputs. Iteratively, the model receives inputs from the training dataset and generates outputs that are compared to the ground truth, and error guides updates to the model parameters. Over time, the model uses the training dataset to learn representations of inputs that result in the desired target outputs.

[0154] The same machine learning technique can produce different models in terms of architecture, weights, and performance depending on which training constraints are chosen. The optimal training constraints, called hyperparameters, are different in each machine learning context and are set before training, as opposed to other model parameters that are developed during training.

[0155] Hyperparameter optimization, also known as tuning, is a step in machine learning development that often has a large impact on a model's ability to generalize successfully. The choice of hyperparameters can lead to a model that overfits or underfits to the same training set. One common approach to explore the space of hyperparameter ranges is a grid search, in which each time a set of hyperparameters is sampled, a model is trained on the training dataset and then evaluated on an independent dataset (the tuning dataset). Performance evaluation on the tuning set guides the selection of optimal training constraints that generate a model that is robust enough to perform beyond the training set.

[0156] Model testing is performed to evaluate the final model fit on the training dataset using the selected hyperparameters. Testing is performed on an independent dataset that was not used for training or tuning to evaluate how well the model generalizes and derives patterns beyond previously encountered data.

[0157] Identifying the Data Set To develop the 3i score model, a comprehensive set of ICI studies (Table 9) with individual patient concentration data was utilized. Studies were selected such that the training and tuning sets included data from patients with different tumor types and across different lines of treatment.

[0158] Separate datasets for training, tuning, and testing The data split for training, tuning and testing was as follows: Training set (used for final input feature selection, model training and hyperparameter optimization): - Study 1108 (N=923), Study 06 (N=353), Study 10 (N=164), Study 21 (N=98) - ATLANTIC (D4191C00003), a phase II, open-label, multicenter, international study of MEDI4736 in patients with locally advanced or metastatic non-small cell lung cancer (stage IIIB-IV) previously treated with at least two prior systemic therapies, including one platinum-based chemotherapy regimen (N=371) - CONDOR (D4193C00003), a multicenter, international, phase II study of MEDI4736 monotherapy, tremelimumab monotherapy and MEDI4736 in combination with tremelimumab in patients with recurrent or metastatic squamous cell carcinoma of the head and neck (N=193); - HAWK (D4193C00001), a multicenter, international, phase II study of MEDI4736 monotherapy in patients with recurrent or metastatic squamous cell carcinoma of the head and neck (N=111). Note: Patients who were censored before 12 weeks were not included in the training set. Tuning set (used in model hyperparameter optimization): - MYSTIC (D419AC00001), a Phase III, randomized, open-label, multicenter, international study comparing MEDI4736 in combination with tremelimumab or MEDI4736 monotherapy with platinum-based standard chemotherapy in the first-line treatment of patients with advanced or metastatic non-small cell lung cancer (N=1118). Test set (used for model testing): - EAGLE (D4193C00002), a Phase III, randomized, open-label, multicenter, international study comparing MEDI4736 monotherapy and MEDI4736 in combination with tremelimumab to standard of care in patients with recurrent or metastatic squamous cell carcinoma of the head and neck (N=736) The total training dataset includes 2213 patients, including the majority of patients diagnosed with NSCLC, SCCHN, UBC and gastric cancer, as summarized in Table 10.

[0159] JPEG2025041587000010.jpg90170

[0160] Feature Selection One strength of using a particular machine learning algorithm is the ability to determine the importance of features and weight predictive features over non-predictive features. Considering that the complexity of the model increases with the number of features or dimensions and that one of the requirements was to use routinely available laboratory measurements (i.e., laboratory measurements covering hematology, chemistry, and thyroid hormones available in the clinical trials selected here), a comprehensive starting list of 37 candidate predictors was collated and summarized in Table 11. It was planned that unimportant or highly correlated features would be removed during the development stage.

[0161] JPEG2025041587000011.jpg215170

[0162] Choosing a Machine Learning Method (Gradient Boosting Trees) Gradient boosting modeling method was chosen to predict 12-week EM. The particular version of gradient boosting method employed is called XGBoost (Extreme Gradient Boosting). This machine learning method is more favored by the statistical community due to its robust predictive performance, its speed (i.e., scalability to parallel processes) and adjustment for missing values ​​(Friedman, et al. 2000; Friedman, et al. 2001; Chen and Guestrin 2016). In gradient boosting, decision trees are used as weak learners. To further improve model fitting, one variable is selected each time to split the node into two subnodes until no split is available. A separate tree is built on the residuals of the previous tree. Variables can be selected multiple times within one tree and in different trees. Decision trees are built successively until a stopping criterion is met. The final gradient boosting model is built by additively assembling decision trees, possibly applying a weighted learning rate to each one. By choosing appropriate model hyperparameters, XGBoost models can be trained to minimize prediction error at controlled model complexity, which leads to a good balance of bias-variance tradeoff. XGBoost models handle missing values ​​by learning a default direction: internally, XGBoost learns the direction to go when values ​​are missing.

[0163] Test Plan A testing plan was designed to evaluate the 3i score by retrospective application to the test set of the ICI clinical trial (EAGLE), which was excluded from training and tuning in order to provide an independent test set on which to evaluate the model in terms of (1) its ability to generalize beyond the training set as defined by true positive rate (TPR), false positive rate (FPR) and partial area under the curve (AUC) in the specificity region of 0.7-0.9, (2) its comparison with other previously developed risk scores, and (3) its clinical utility as defined by the increase in treatment effect after excluding the 3i score high-risk group.

[0164] Code testing was not applicable in the test plan as no new functionality was introduced, nor were any existing functionality extended to the open source R packages used for development.

[0165] development Development environment All development was performed with R v.3.5.1 on a x86_64 pc linux gnu (64-bit) operating system. All open source packages used in development are listed in Table 12.

[0166] JPEG2025041587000012.jpg191170

[0167] Separating and labeling the dataset The resulting dataset was preprocessed internally according to internal data management standards. Patients who died within 12 weeks (i.e., time from randomization / first dose to death ≤ 12 weeks) were labeled as EM. Patients who survived beyond 12 weeks (i.e., time from randomization / first dose to death > 12 weeks) were not labeled as EM. Patients who were censored within 12 weeks (i.e., patients lost to follow-up at ≤ 12 weeks) were not included in the training set during model development. No other data manipulation was performed.

[0168] Feature Selection Feature selection was performed to explore a wide range of candidate predictors for the 3i score model and return a subset of predictors for which a performance model could be fitted. First, a full model was fitted to the training data (Figure 52) with 37 predictors using a gradient boosting fitting procedure (see Table 11). Model fitting was performed using 10-fold cross-validation. A grid search of hyperparameters was performed and the model that gave the highest partial AUC was used to guide feature selection.

[0169] A gradient boosting model was fitted to the training data to calculate variable importance. Variable importance is quantified by the extent to which the inclusion of the variable resulted in improved accuracy. Initially, all available pretreatment blood test variables (n=31) covering hematology, clinical chemistry, and thyroid hormones were included in this first step (i.e., all clinical test variables). Clinical variables well recognized to be associated with prognosis in oncology such as age ("AGE"), sex ("SEX"), baseline tumor size (estimated by RECIST, "TUMSZINV"), body mass index ("BMI"), tumor type ("typ"), and presence of liver metastases ("LIVERBL") were also introduced into the model. As shown in Figure 53, albumin, LDH, neutrophils, neutrophil / leukocyte ratio (NEUTLE), NLR, gamma-glutamyltransferase (GGT), and aspartate aminotransferase (AST) were found to be the most important variables. Neutrophils and NEUTLE were highly correlated (r=0.98, p<0.001), so NEUTLE was not selected to build the final model. Based on this feature importance analysis and subject matter expert input, the final set of input features to train and tune the 3i score model was albumin, LDH, neutrophils, NLR, GGT, and AST.

[0170] Biological relevance of blood test variables selected for the 3i score model All six blood test variables assessed by standard laboratory criteria and identified by feature selection have been previously reported to be prognostic in a variety of tumor types.

[0171] High NLR is associated with poor prognosis in many tumor types, including gastrointestinal (Bowen, et al. 2017), melanoma (Ding, et al. 2018), NSCLC (Fukui, et al. 2019), UC (Sacdalan, et al. 2018) and HNSCC (Yu, et al. 2018, Tham, et al. 2018). This ratio likely reflects an assessment metric of immune fitness that may affect tumor immune responses, and the association with poor prognosis is determined to be independent of the type of ICI. Fukui et al. observed a significant association between high NLR and poor prognosis in NSCLC patients treated with nivolumab. In a meta-analysis of ICI studies in melanoma, NSCLC and UC, Sacdalan et al. observed an association between high NLR and poor outcome, suggesting its potential as a prognostic marker. It has also been speculated that tumor burden and chronic inflammation may contribute to high NLR and subsequently poor prognosis (Bigot, et al. 2017).

[0172] Absolute neutrophil count (ANC) is a surrogate marker of physiological stress, including inflammation, and in patients with NSCLC, high ANC after nivolumab treatment has been observed to be independently associated with poor overall survival (Khunger, et al. 2018).

[0173] Albumin (ALB) is a well-known marker of malnutrition and cachexia in cancer patients, and low ALB has been observed to be associated with poor overall survival in many malignancies, including SCCHN, breast cancer, lung cancer, and gastrointestinal cancer (Gupta, et al. 2010).

[0174] LDH is an established marker reflecting enhanced cell turnover primarily in malignant disease, but also in non-malignant diseases (e.g., hemolytic anemia). A recent FDA analysis of early deaths in ICI studies reported elevated LDH as a significant risk factor primarily in melanoma patients (Mulkey, et al. 2019). In a systematic review of 76 studies, high LDH was associated with poor prognosis in renal, melanoma, gastric, prostate, nasopharyngeal and lung cancers (Petrelli, et al. 2015). In advanced cancers, elevated circulating LDH concentrations reflect increased metabolic activity and rapid cell proliferation that results in increased leakage of the intracellular enzyme LDH into the blood circulation.

[0175] In cancer patients, abnormalities in liver function tests such as AST and GGT are closely monitored. Abnormalities in liver function tests not only reflect existing disease, but also cancer progression / liver metastasis and drug toxicity, all of which are known causes of death in cancer patients. Thus, AST and GGT have been reported as prognostic markers in cancer (Freis, et al. 2017; Luo, et al. 2017), and GGT has been suggested to play a role in tumor initiation, invasion, and drug resistance.

[0176] 3iScore model training and hyperparameter selection Only the most predictive biological variables (NLR, neutrophils, albumin, LDH, AST, GGT) and tumor type, previously described in the feature selection step (Figure 53), were retained for training the final 3i score model (Figure 52). XGBoost has 12 tree booster hyperparameters that set training constraints such as base learner type, learning rate, maximum tree depth allowed, and weight regularization. Since each set of hyperparameters produces a different XGBoost model during training, the MYSTIC data was used as a tuning set to guide the selection of the optimal set of hyperparameters. A grid search was performed for max.depth, gamma, min.child.weight, colsample.bytree, and subsample hyperparameters. A pool of XGBoost models was generated by fitting models to the training dataset using different sets of hyperparameters. Model performance on the training set was quantified as true positive rate and partial AUC (specificity 0.7-0.9). Hazard ratios and p-values ​​(Cox regression) for the interaction between 3i score status and treatment in the PD-L1 population were calculated and used as a quantification of model performance in the tuning set (MYSTIC). The model with the smallest interaction p-value in the tuning set (MYSTIC) and high partial AUC in the training set was selected as the final model.

[0177] The final 3i-score model, selected from the pool of 432 models, had a pAUC (0.7, 0.9) of 0.83 (ranked 8th among all 432 models) on the training dataset and was trained with the following non-default XGBoost hyperparameters: Maximum tree depth (max.depth): 3 Minimum loss reduction (gamma) required to create an additional partition at the leaf nodes of the tree: 1 Minimum sum of instance weights (hessian) required for children (min.child.weight): 5 Subsample ratio of columns when building each tree (colsample.bytree): 1 Subsample ratio for training instances: 0.6 · scale_pos_weight: negative_examples / positive_examples Rounds:22

[0178] To describe the properties of the final model, the importance of each variable selected in the final model was calculated. As shown in Figure 54, NLR, NEUT, ALB, LDH and GGT are classified as more important than AST and tumor type.

[0179] Final 3i score model The following six significant predictors were retained in the final model: NLR, NEUT, ALB, LDH, GGT, AST, and tumor type. The model generates a score (i.e., a value between 0 and 1) that represents the probability of death occurring at ≦12 weeks for each patient. The 3i score is then converted into a state that assigns the patient to a high or low prognosis or risk category. The cutoff was determined to allow for a 10% false positive rate in predicting EM in the training dataset and was calculated to be 0.649 on July 2, 2019. Patients above the cutoff value of 0.649 are identified as high risk of EM, and patients below the cutoff value (0.649) are identified as low risk of EM. For patients lacking any of the six laboratory values, the 3i score is not calculated. This decision was supported by the expectation that globally, patients would have readily available data on these variables, as these are standard laboratory criteria that are routinely collected.

[0180] 3i vs. GRIm and LIPI on training and tuning datasets When applied to the training set (N=2213 patients), the TPR at 12 weeks was 67% and the FPR was 10%. The performance as defined by pAUC(0.7, 0.9) was 0.83. Patients identified as having high-risk 3i score status (486 / 2213 patients) had a median OS of 9.29 weeks (95%CI: 8.29 weeks, 9.86 weeks) and patients with low-risk 3i score status had a median OS of 61.43 weeks (95%CI: 57.71, 66.14) (see Figure 55).

[0181] Patients identified as high risk by the 3i score (see Figures 56A and 56B) had a shorter median OS compared to patients identified as high risk by other prognostic models developed to predict early death in patients with advanced or metastatic cancer: GRIm (624 patients, median OS 14.29 weeks [95% CI: 12.71, 15.71]); and LIPI score (363 patients, median OS 14.00 weeks [95% CI: 11.43, 16.00]). ULN=240 is used for LDH to calculate the GRIm score and LIPI score.

[0182] When applied to the tuning set (MYSTIC) (N=1118) (see Figures 57A and 57B), durvalumab-treated patients had a 12-week TPR of 52% and an FPR of 11%. The performance as defined by pAUC(0.7, 0.9) was 0.63. Patients identified as having high-risk 3i score status (69 / 374 patients) had a median OS of 12.00 weeks (95%CI: 9.57, 22.00) and patients with low-risk 3i score status (289 / 374 patients) had a median OS of 68.71 weeks (95%CI: 60.29, 81.00). Patients identified as high risk by the 3i score had a shorter median OS compared with patients identified as high risk by other prognostic models: GRIm score (115 patients, median OS 20.43 [95% CI: 14.29, 28.71]); and LIPI score (71 patients, median OS 18.14 [95% CI: 10.00, 33.29]).

[0183] In the MYSTIC intent-to-treat population (ITT) (see Figures 58A-58C), the median OS in the durvalumab and SoC groups was 53.29 and 51.14 weeks, respectively. When applying the 3i score to the MYSTIC ITT population (N=746), 120 patients (16%) were identified as high risk, of which 51 patients were in the SoC group and 69 patients were in the durvalumab group. The median OS of patients identified as having high-risk 3i score status (120 / 746, 16% of patients) appeared to show poorer outcomes in the durvalumab group (median OS: 12.00 weeks) than in the chemotherapy group (median OS: 29.14 weeks, HR=1.39 [95%CI: 0.947, 2.053]). In low-risk patients, there was still only a small improvement in treatment efficacy (HR=0.89 [95%CI:0.735, 1.079]) and crossing of the OS curves was not excluded, possibly because PD-L1 low / negative patients may not benefit from monotherapy ICIs compared with active chemotherapy comparators for advanced disease.

[0184] In the MYSTIC PD-L1 ≥ 25% subgroup (see Figures 59A-59C), 56 patients (17%) were identified as 3i score high risk, 24 in the SoC group and 32 in the durvalumab group. Similar to ITT, in 3i score high risk patients in the PD-L1 ≥ 25% subgroup, the median OS in the durvalumab group was shorter than that in the chemotherapy group (chemotherapy 35.79 weeks vs. durvalumab 11.36 weeks, HR = 1.533 [95% CI: 0.873, 2.726], Figure 59C). Importantly, excluding patients with 3i score high risk status (56 / 325 patients, 17% of patients excluded by 3i score) from the MYSTIC PD-L1 ≥ 25% subgroup reduced the crossing of OS curves previously observed (Figure 59B). The treatment effect of durvalumab on chemotherapy was greater (HR: 0.626 [95% CI: 0.459, 0.849]) compared with the PD-L1 ≥ 25% population in the initial primary analysis (HR: 0.76 [95% CI: 0.59, 0.98]).

[0185] 3i Score Test Application of 3i score model on EAGLE (test set) When applied to the test set (EAGLE) (N=736) (see Figures 60A and 60B), the 12-week TPR for durvalumab-treated patients was 57% and the FPR was 16%. The performance as defined by pAUC(0.7, 0.9) was 0.64. Patients identified as having high-risk 3i score status (59 / 240 patients) had a median OS of 12.14 weeks (95%CI: 8.57, 19.14) and patients with low-risk 3i score status (172 / 240 patients) had a median OS of 48.00 weeks (95%CI: 41.29, 58.71). Patients identified as high risk by the 3i score had shorter OS compared with patients identified as high risk by other prognostic models: GRIm score (67 patients, median OS 15.71 weeks [95% CI: 12.14, 21.86]); and LIPI score (38 patients, median OS 19.36 weeks [95% CI: 13.71, 35.43]).

[0186] In EAGLE ITT (see Figures 61A-61C), the median OS in the durvalumab group (N=240) and the SoC group (N=249) was 33.14 weeks and 36.00 weeks, respectively. When applying the 3i score in EAGLE ITT, a total of 121 patients (25%) were identified as high risk, of which 62 were in the SoC group and 59 in the durvalumab group. The median OS of patients identified as high risk by the 3i score appeared to be worse in the durvalumab group (median OS=12.14 weeks) than in the chemotherapy group (median OS=16.00 weeks, HR 1.30 [95% CI: 0.86, 1.99]). When excluding patients identified as high risk by the 3i score from EAGLE ITT (i.e., 25% of patients), the crossover of the overall survival curves was reduced (Figure 61B). The magnitude of the treatment effect of durvalumab was increased (HR: 0.75 [95% CI: 0.58, 0.96]) compared with the original ITT population (HR: 0.88 [95% CI: 0.72, 1.08]).

[0187] In the PD-L1 ≥ 25% subgroup (N = 140) (see Figures 62A-62C), a total of 43 patients (31%) were identified as high risk by the 3i score, 20 in the SoC group and 23 in the durvalumab group (Figure 62C). Similar to what was seen in the MYSTIC dataset (tuning set), excluding these patients partially eliminated the crossover of the overall survival curves (Figure 62B). The magnitude of the treatment effect of durvalumab was increased (HR: 0.66 [95% CI: 0.38, 1.12]) compared to the original primary analysis PD-L1 ≥ 25% population (HR: 0.93 [95% CI: 0.63, 1.39]).

[0188] Distribution of laboratory tests in 3i score low-risk patients versus 3i score high-risk patients In the MYSTIC and EAGLE studies, Wilcoxon tests were used to compare the six laboratory values ​​used in the 3i score between patients identified as high risk (predicted life expectancy ≦12 weeks) or low risk (predicted life expectancy >12 weeks) by the 3i score. For all six parameters, values ​​from patients with high and low risk 3i score status were significantly different (p<0.05) (see Figures 63 and 64).

[0189] Example 11 –Determination of patient immune fitness overview The variables included in the 3i score are associated with host immune fitness and cancer-associated inflammation. Furthermore, additional translational analyses performed on samples obtained in clinical trials show that 3i high-risk patients are characterized by high baseline concentrations of peripheral blood immunosuppressive cells (MDSCs), high expression of immunosuppressive gene signatures and proinflammatory protein concentrations in serum (CRP, IL6 and IL8). Interestingly, unlike immune checkpoint blockade, chemotherapy is well recognized to reduce IL6, IL8 and CRP concentrations. This finding could explain the phenomenon of early death observed with ICB versus chemotherapy in patients with advanced / metastatic cancer.

[0190] To investigate whether the 3i score can be used as an indicator of patient immune fitness, the baseline molecular profile of 3i high-risk patients was assessed by blood gene expression, peripheral blood immune profiling by flow cytometry and serum protein expression.

[0191] result Peripheral blood gene expression 3i-high (poor prognosis) is associated with lower baseline T effector expression signature and higher immunosuppressive cell gene expression signature compared to 3i-low in NSCLC, HNSCC and UBC patients, regardless of the patient's PD-L1 status (see Figures 65 and 66). Across MYSTIC, ATLANTIC, CP1108, HAWK, CONDOR, 006 and 10 studies, 51 / 80 genes measured by qPCR were found to be differentially expressed between 3i-high and 3i-low with a cutoff of 1.5-fold difference and p-value <0.01% (Wilcoxon test).

[0192] Peripheral immune profiling by flow cytometry High 3i was associated with lower baseline T, NK and B cells and higher absolute numbers of monocytic myeloid-derived suppressor cells (MNC MDSC) compared to low 3i in NSCLC patients (see Figures 67 and 68). 11 / 22 immunophenotypes from five flow cytometry panels evaluated in CP1108, 006 and Mystic NSCLC patients were found to be differentially expressed between 3i high and 3i low patient groups (>20% CV with Wilcoxon test p-value <0.05%).

[0193] Serum protein expression High 3i (poor prognosis) was significantly associated with higher baseline proinflammatory protein concentrations, especially C-reactive protein (CRP) and interleukins IL6 and IL8, compared to 3i-low NSCLC, HNSCC and UBC patients (see Figure 69). Across the CP1108, HAWK and CONDOR studies, 16-22 / 66 serum protein concentrations assessed by Luminex immunoassay (Myriad RBM) were found to be differentially expressed in 3i-high compared to 3i-low NSCLC, HNSCC and UBC patients (% difference compared to 3i-low >20%CV with Wilcoxon test p-value <0.05%). In addition to CRP, IL6 and IL8, pro-vascular and angiogenic proteins (e.g. ANGPT2, VEGF, ICAM1 and vWF) were found to be higher in the 3i-high patient group compared to the 3i-low patient group.

[0194] conclusion The 3i index may have clinical utility in identifying patients who may lack immune fitness. Lack of immune fitness suggests that such patients should postpone ICI treatment until an effective treatment is available to improve immune fitness, or treat the patient with a non-ICI therapeutic. Thus, therapeutic interventions that enhance immune system fitness in 3i-high patients may allow these patients to receive ICI treatment with reduced EM risk, further increasing the chances of therapeutic benefit from ICI treatment.

[0195] Examples of Computing Systems 70 is a simplified block diagram illustrating a computing device 7000 showing some of the components that may be included in a computing device configured to operate according to the embodiments described herein. The computing device 7000 may be a client device (e.g., a device that is actively operated by a user), a server device (e.g., a device that provides computing services to client devices), or some other type of computing platform. Some server devices may act as client devices at times to perform certain operations, and some client devices may have server functionality built in.

[0196] In this example, computing device 7000 includes a processor 7002, memory 7004, a network interface 7006, and input / output units 7008, all of which may be coupled by a system bus 7010 or similar mechanism. In some embodiments, computing device 7000 may include other components and / or peripheral devices (e.g., removable storage, printers, etc.).

[0197] The processor 7002 may be one or more of any type of computer processing element, such as a central processing unit (CPU), a coprocessor (e.g., a mathematical, graphics or encryption coprocessor), a digital signal processor (DSP), a network processor and / or in the form of an integrated circuit or controller that performs processor operations. In some cases, the processor 7002 may be one or more single-core processors. In other cases, the processor 7002 may be one or more multi-core processors having multiple independent processing units. The processor 7002 may also include a register memory for temporarily storing instructions and associated data being executed, and a cache memory for temporarily storing recently used instructions and data.

[0198] Memory 7004 may be any form of computer usable memory, including, but not limited to, random access memory (RAM), read only memory (ROM), and non-volatile memory (e.g., flash memory, hard disk drives, solid state drives, compact discs (CDs), digital video discs (DVDs), and / or tape storage). Memory 7004 thus represents both primary and long term storage. Other types of memory may include biomedical memory.

[0199] The memory 7004 may store program instructions and / or data on which the program instructions may operate. By way of example, the memory 7004 may store these program instructions on a non-transitory computer-readable medium such that the instructions are executable by the processor 7002 to perform any of the methods, processes, or operations disclosed herein or in the accompanying drawings.

[0200] As shown in FIG. 70, memory 7004 may include firmware 7004A, kernel 7004B, and / or applications 7004C. Firmware 7004A may be program code used to boot or start part or all of computing device 7000. Kernel 7004B may be an operating system that includes modules for memory management, process scheduling and management, input / output, and communications. Kernel 7004B may also include device drivers that allow the operating system to communicate with hardware modules (e.g., storage devices, network interfaces, ports, and buses) of computing device 7000. Applications 7004C may be one or more user space software programs, such as a web browser or email client, as well as any software libraries used by these programs. Memory 7004 may also store data used by these and other programs and applications.

[0201] The network interface 7006 may take the form of one or more wired interfaces, such as Ethernet (e.g., Fast Ethernet, Gigabit Ethernet, etc.). The network interface 7006 may also support communication over one or more non-Ethernet media, such as coaxial cable or power line, or over wide area media, such as Synchronous Optical Network (SONET) or Digital Subscriber Line (DSL) technology. The network interface 7006 may further take the form of one or more wireless interfaces, such as IEEE 802.11 (Wifi), BLUETOOTH, Global Positioning System (GPS), or wide area radio interfaces. However, other forms of physical layer interfaces and other types of standard or proprietary communication protocols may be used in the network interface 7006. Furthermore, the network interface 7006 may include multiple physical interfaces. For example, some embodiments of the computing device 7000 may include Ethernet, BLUETOOTH, and Wifi interfaces.

[0202] The input / output unit 7008 can facilitate user and peripheral device interaction with the computing device 7000. The input / output unit 7008 can include one or more types of input devices, such as a keyboard, a mouse, a touch screen, etc. Similarly, the input / output unit 7008 can include one or more types of output devices, such as a screen, a monitor, a printer, and / or one or more light emitting diodes (LEDs). Additionally or alternatively, the computing device 7000 can communicate with other devices using, for example, a Universal Serial Bus (USB) or a High Definition Multimedia Interface (HDMI) port interface.

[0203] In some embodiments, one or more computing devices, such as computing device 7000, may be deployed to support the machine learning architecture. The exact physical location, connectivity, and configuration of these computing devices may be unknown and / or unimportant to the client device. Thus, the computing devices may be referred to as "cloud-based" devices that may be housed in a variety of remote data center locations.

[0204] Figure 71 illustrates a cloud-based server cluster 7100 according to an example embodiment. In Figure 71, the operations of a computing device (e.g., computing device 7000) may be distributed among server devices 7102, data storage 7104, and routers 7106, all of which may be connected by a local cluster network 7108. The number of server devices 7102, data storage 7104, and routers 7106 in the server cluster 7100 may depend on the computing tasks and / or applications assigned to the server cluster 7100.

[0205] For example, a server device 7102 may be configured to perform various computing tasks of the computer device 7000. Thus, the computing tasks may be distributed among one or more server devices 7102. To the extent that these computing tasks can be performed in parallel, such distribution of tasks may reduce the total time to complete these tasks and return results. For simplicity, both the server cluster 7100 and the individual server devices 7102 may be referred to as "server devices." This nomenclature should be understood to mean that one or more separate server devices, data storage devices, and cluster routers may be involved in the operation of a server device.

[0206] The data storage 7104 may be a data storage array including a drive array controller configured to manage read and write access to a group of hard disk drives and / or solid state drives. The drive array controller, alone or in combination with the server device 7102, may be configured to manage backup or redundant copies of data stored in the data storage 7104 to protect against drive failure or other types of failures that prevent one or more server devices 7102 from accessing units of the data storage 7104. Other types of memory besides drives may also be used.

[0207] The router 7106 may include a network device configured to provide internal and external communications for the server cluster 7100. For example, the router 7106 may include one or more packet switching and / or routing devices (including switches and / or gateways) configured to provide (i) network communications between the server devices 7102 and the data storage 7104 via a local cluster network 7108, and / or (ii) network communications between the server cluster 7100 and other devices via a communication link 7110 to a network 7112.

[0208] Further, the configuration of the router 7106 may be based, at least in part, on the data communication requirements of the server device 7102 and data storage 7104, the latency and throughput of the local cluster network 7108, the latency, throughput and cost of the communication links 7110, and / or other factors that may contribute to the cost, speed, fault tolerance, resilience, efficiency and / or other design goals of the system architecture.

[0209] As a possible example, data storage 7104 may include any type of database, such as a Structured Query Language (SQL) database. Various types of data structures may store information in such a database, including, but not limited to, tables, arrays, lists, trees, and tuples. Additionally, any database in data storage 7104 may be monolithic or distributed across multiple physical devices.

[0210] The server device 7102 may be configured to send and receive data to and from the data storage 7104. This sending and retrieval may take the form of SQL queries or other types of database queries and the output of such queries, respectively. Additional text, images, video, and / or audio may also be included. Additionally, the server device 7102 may organize the received data into a web page or web application representation. Such a representation may take the form of a markup language, such as Hypertext Markup Language (HTML), Extensible Markup Language (XML), or other standardized or proprietary formats. Additionally, the server device 7102 may have the capability to execute various types of computerized scripting languages, such as, but not limited to, Perl, Python, PHP Hypertext Preprocessor, Active Server Pages (ASP), JAVASCRIPT®, and the like. Computer program code written in these languages ​​may facilitate the provision of web pages to client devices and the client device's interaction with the web pages. Alternatively or in addition, JAVA® may be used to facilitate the generation of web pages and / or provide web application functionality.

[0211] Example of operation 72 and 73 are flow charts illustrating exemplary embodiments. The processes illustrated by FIGS. 72 and 73 may be performed by a computing device, such as computing device 7000, and / or a cluster of computing devices, such as server cluster 7100. However, the processes may also be performed by other types of devices or device subsystems. The embodiments of FIGS. 72 and 73 may be simplified by omitting any one or more of the features shown therein. Furthermore, these embodiments may be combined with each other's features, aspects and / or implementations and / or any of those described in the previous figures or otherwise herein.

[0212] Block 7200 of Figure 72 may include obtaining model training data, which includes a first set of cancer patients undergoing cancer treatment, and which correlates (i) results from clinical tests performed on the first set of cancer patients with tumor types of the first set of cancer patients, and (ii) whether each patient from the first set of cancer patients died within a reference number of weeks from the start of cancer treatment. The clinical tests may include tissue biopsies, cerebrospinal fluid, lymphatic fluid, whole blood, serum, blood cells, urine, sweat, tears, saliva, and / or feces tests, or any type of test that indicates clinical pathological markers.

[0213] Block 7202 includes training a series of decision trees based on the model training data until a stopping condition is reached, where each successive decision tree in the series is constructed based on the residual values ​​of its predecessor decision tree in the series.

[0214] Block 7204 includes creating a gradient boosting machine learning model as an additive function of the series of decision trees, where the gradient boosting machine learning model predicts whether the further cancer patient will die within a reference number of weeks based on the additive function applied to results from clinical tests performed on the further cancer patient and the tumor type of the further cancer patient.

[0215] In some embodiments, the baseline number of weeks is 12 weeks, although other values ​​are possible (e.g., 8 weeks, 10 weeks, 15 weeks, or 20 weeks).

[0216] In some embodiments, results from the laboratory tests provide one or more of the following: neutrophil to lymphocyte ratio, neutrophil concentration, albumin concentration, lactate dehydrogenase concentration, aspartate aminotransferase concentration, or gamma-glutamyltransferase concentration. Any or all of these results can be used.

[0217] In some embodiments, the neutrophil-to-lymphocyte ratio, neutrophil concentration, albumin concentration, lactate dehydrogenase concentration, aspartate aminotransferase concentration, and gamma-glutamyltransferase concentration are selected from 20 or more clinical tests performed on a first set of cancer patients. The selecting may include (i) training a gradient boosting machine learning model on the 20 or more clinical tests, (ii) for each clinical test, training respective variations of the gradient boosting machine learning model on the 20 or more clinical tests, excluding the respective clinical test; and (iii) determining a predictive value for each of the 20 or more clinical tests by comparing the accuracy of the variations of the gradient boosting machine learning model. The 20 or more clinical tests may be selected from the predictors in Table 11 and may include non-test predictors such as gender and age.

[0218] In some embodiments, the tumor type is one of the following: non-small cell lung cancer, urothelial carcinoma, squamous cell carcinoma of the head and neck, gastric carcinoma, renal cell carcinoma, or Hodgkin's lymphoma. Other types are possible.

[0219] In some embodiments, the additive function applies a multiplicative learning rate to the outcome from each of a series of decision trees.

[0220] In some embodiments, predicting whether the additional cancer patient will die within the reference number of weeks comprises determining the probability that the additional cancer patient will die within the reference number of weeks by using an additive function.

[0221] In some embodiments, predicting whether the additional cancer patients will die within the reference number of weeks further comprises placing the additional cancer patients in either a high or low risk of death category based on the probability and the cutoff value. In some embodiments, the cutoff value is selected such that there is no false positive rate of more than 5-20% when predicting whether the first set of cancer patients will die within the reference number of weeks. Optionally, the cutoff value may be selected such that the false positive rate is 10%.

[0222] In some embodiments, cancer patients who fall into a high risk of death category are excluded from clinical trials for the treatment of cancer, and cancer patients who fall into a low risk of death category are included in clinical trials for the treatment of cancer.

[0223] In some embodiments, the stopping condition is that the set of decision trees includes a reference number of decision trees. Alternatively or additionally, the stopping condition is that the residual values ​​of a particular decision tree in the set of decision trees are all less than a reference value.

[0224] Some embodiments may further include obtaining model tuning data, the model tuning data including a second set of cancer patients receiving cancer treatment, the model tuning data correlating (i) results from a clinical test performed on the second set of cancer patients with the tumor type of the second set of cancer patients and (ii) whether the second set of cancer patients died within a reference number of weeks from initiation of cancer treatment; creating two or more variations of a gradient boosting machine learning model by training a series of decision trees for each variation having a different set of hyperparameters; applying the variations to the model training data and the model tuning data; and selecting from the variations as the gradient boosting machine learning model a particular variation that provides within a reference accuracy a prediction of whether the first set of cancer patients and the second set of cancer patients will die within a reference number of weeks from initiation of cancer treatment.

[0225] The accuracy of the reference can be based on the true positive rate or true negative rate of predicting whether a first set of cancer patients will die within a reference number of weeks from the start of cancer treatment. Alternatively, or in addition, the accuracy of the reference can be based on the p-value of the interaction between (i) the prediction whether a second set of cancer patients will die within a reference number of weeks from the start of cancer treatment and (ii) the PD-L1 positive patients in the second set of cancer patients.

[0226] In some embodiments, the hyperparameters include one or more of the following: maximum depth of a decision tree, set to 2, 3, 4, 5, or 6; minimum loss reduction required to split a node in a decision tree, set to 0, 1, 2, 3, or 4; minimum sum of weights of all observations required at a child node in a decision tree, set to 3.0 to 7.0; fraction of input parameters randomly sampled for each of the decision trees, set to 0.7 to 1.0; fraction of observations randomly sampled for each of the decision trees, set to 0.4 to 0.8; number of decision trees in the set of decision trees, set to 10 to 25. Other ranges may be used as well.

[0227] Block 7300 of FIG. 73 may include obtaining a gradient boosting machine learning model, the gradient boosting machine learning model being trained with model training data, the model training data relating (i) results from clinical tests performed on a first set of cancer patients and tumor types of the first set of cancer patients with (ii) whether each patient from the first set of cancer patients died within a reference number of weeks from the start of cancer treatment, and training the gradient boosting machine learning model included (i) training a series of decision trees until a stopping condition is reached, each successive decision tree in the series being constructed based on residual values ​​of its predecessor in the series, and (ii) determining an additive function of the series of decision trees. The clinical tests may include tissue biopsies, cerebrospinal fluid, lymph, whole blood, serum, blood cells, urine, sweat, tears, saliva, and / or feces tests, or any type of test indicative of clinical pathological markers.

[0228] Block 7302 may include obtaining results from laboratory tests performed on additional cancer patients.

[0229] Block 7304 may include obtaining tumor types for additional cancer patients.

[0230] Block 7306 may include applying the gradient boosting machine learning model to results from the clinical tests performed on the additional cancer patients and the tumor types of the additional cancer patients.

[0231] Block 7308 may include receiving a prediction from the gradient boosting machine learning model of whether an additional cancer patient will die within a reference number of weeks.

[0232] In some embodiments, the baseline number of weeks is 12 weeks, although other values ​​are possible (e.g., 8 weeks, 10 weeks, 15 weeks, or 20 weeks).

[0233] In some embodiments, results from the laboratory tests provide one or more of the following: neutrophil to lymphocyte ratio, neutrophil concentration, albumin concentration, lactate dehydrogenase concentration, aspartate aminotransferase concentration, or gamma-glutamyltransferase concentration. Any or all of these results can be used.

[0234] In some embodiments, the tumor type is one of the following: non-small cell lung cancer, urothelial carcinoma, squamous cell carcinoma of the head and neck, gastric carcinoma, renal cell carcinoma, or Hodgkin's lymphoma. Other types are possible.

[0235] In some embodiments, the additive function applies a multiplicative learning rate to the outcome from each of a series of decision trees.

[0236] In some embodiments, predicting whether the additional cancer patient will die within the reference number of weeks comprises determining the probability that the additional cancer patient will die within the reference number of weeks by using an additive function.

[0237] In some embodiments, predicting whether the additional cancer patients will die within the reference number of weeks further comprises placing the additional cancer patients into either a high or low risk of death category based on the probability and the cutoff value. In some embodiments, the cutoff value is selected such that there is no false positive rate of more than 5-20% when predicting whether the first set of cancer patients will die within the reference number of weeks. Optionally, the cutoff value may be selected such that the false positive rate is 10%.

[0238] In some embodiments, cancer patients who fall into a high risk of death category are excluded from clinical trials for the treatment of cancer, and cancer patients who fall into a low risk of death category are included in clinical trials for the treatment of cancer.

[0239] In some embodiments, the stopping condition is that the set of decision trees includes a reference number of decision trees. Alternatively or additionally, the stopping condition is that the residual values ​​of a particular decision tree in the set of decision trees are all less than a reference value.

[0240] In some embodiments, the gradient boosting machine learning model has also been trained using model tuning data, the model tuning data including a second set of cancer patients receiving cancer treatment, the model tuning data associating (i) results from clinical tests performed on the second set of cancer patients and tumor types of the second set of cancer patients with (ii) whether the second set of cancer patients died within a reference number of weeks from initiation of cancer treatment, and training the gradient boosting machine learning model using the model tuning data includes (i) generating two or more variations of the gradient boosting machine learning model by training a series of decision trees for each variation having a different set of hyperparameters; (ii) applying the variations to the model training data and the model tuning data; and (iii) selecting from the variations as the gradient boosting machine learning model a particular variation that provides a prediction within a reference accuracy of whether the first set of cancer patients and the second set of cancer patients will die within a reference number of weeks from initiation of cancer treatment.

[0241] The accuracy of the reference can be based on the true positive rate or true negative rate of predicting whether a first set of cancer patients will die within a reference number of weeks from the start of cancer treatment. Alternatively, or in addition, the accuracy of the reference can be based on the p-value of the interaction between (i) the prediction whether a second set of cancer patients will die within a reference number of weeks from the start of cancer treatment and (ii) the PD-L1 positive patients in the second set of cancer patients.

[0242] In some embodiments, the hyperparameters include one or more of the following: maximum depth of a decision tree, set to 2, 3, 4, 5, or 6; minimum loss reduction required to split a node in a decision tree, set to 0, 1, 2, 3, or 4; minimum sum of weights of all observations required at a child node in a decision tree, set to 3.0 to 7.0; fraction of input parameters randomly sampled for each of the decision trees, set to 0.7 to 1.0; fraction of observations randomly sampled for each of the decision trees, set to 0.4 to 0.8; number of decision trees in the set of decision trees, set to 10 to 25. Other ranges may be used as well.

[0243] conclusion A novel model was developed that relies on baseline criteria, as opposed to subjective clinical criteria such as the ECOG score. Based on evaluation of this index in the MYSTIC and EAGLE trials, it appears to have clinical utility in identifying patients who may have a relatively favorable benefit:risk profile compared to the ITT population.

[0244] All patents and publications mentioned in this specification are herein incorporated by reference to the same extent as if each individual patent and publication was specifically and individually indicated to be incorporated by reference. Citation or identification of any reference in any section of this application shall not be construed as an admission that such reference is available as prior art to the present invention.

[0245] References Arkenau HT, Barriuso J, Olmos D, Ang JE, de Bono J, Judson I, et al.Prospective validation of a prognostic score to improve patient selection for oncology phase I trials.J Clin Oncol.2009;27(16):2692-6. Bellmunt J,de Wit R,Vaughn DJ,Fradet Y,Lee JL,Fong L et al.Pembrolizumab as Second-Line Therapy for Advanced Urothelial Carcinoma.N Engl J Med.2017;376(11):1015-26. Bigot F,Castanon E,Baldini C,Hollebecque A,Carmona A,Postel-Vinay S,et al.Prospective validation of a prognostic score for patients in immunotherapy phase I trials:The Gustave Roussy Immune Score(GRIm-Score).Eur J Cancer.2017;84:212-8. Borghaei H,Paz-Ares L,Horn L,Spigel DR,Steins M,Ready NE,et al.Nivolumab versus Docetaxel in Advanced Nonsquamous Non-Small-Cell Lung Cancer.N Engl J Med.2015;373(17):1627-39. Bowen RC,Little NAB,Harmer JR,Ma J,Mirabelli LG,Roller KD,et al.Neutrophil-to-lymphocyte ratio as prognostic indicator in gastrointestinal cancers:a systematic review and meta-analysis.Oncotarget.2017;8(19):32171-89. Caliskan,A,Bryson JJ,and Narayanan A.Semantics derived automatically from language corpora contain human-like biases.Science.356(6334):183-186,2017. Champiat S,Ferrara R,Massard C,Besse B,Marabelle A,Soria JC,et al.Hyperprogressive disease:recognizing a novel pattern to improve patient management.Nat Rev Clin Oncol.2018;15(12):748-62. Chen T,Guestrin C.XGBoost:A Scalable Tree Boosting System.Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining(KDD).2016;785-794.DOI:http: / / dx.doi.org / 10.1145 / 2939672.2939785. Chen T,He T,Benesty M,Khotilovich V,Tang Y,Cho H,Chen K,Mitchell R,Cano I,Shou T,Li M,Xie J,Lin M,Geng Y,Li Y.R Package’xgboost’,March 12,2019.https: / / cran.r-project.org / web / packages / xgboost / xgboost.pdf. Ding Y,Zhang S,Qiao J.Prognostic value of neutrophil-to-lymphocyte ratio in melanoma:Evidence from a PRISMA-compliant meta-analysis.Medicine(Baltimore).2018;97(30):e11446. Ferris RL,Blumenschein G,Jr.,Fayette J,Guigay J,Colevas AD,Licitra L,et al.Nivolumab for Recurrent Squamous-Cell Carcinoma of the Head and Neck.N Engl J Med.2016;375(19):1856-67. Friedman J,Hastie T,Tibshirani R.(2000).Additive logistic regression:a statistical view of boosting(with discussion and a rejoinder by the authors).The annals of statistics.2000;28(2),337-407. Friedman JH.Greedy function approximation:a gradient boosting machine.Annals of Statistics,pp.2001;1189-1232. Freis P,Graillot E,Rousset P,Hervieu V,Chardon L,Lombard-Bohas C,et al.Prognostic factors in neuroendocrine carcinoma:biological markers are more useful than histomorphological markers.Sci Rep.2017;7:40609. Fukui T,Okuma Y,Nakahara Y,Otani S,Igawa S,Katagiri M,et al.Activity of Nivolumab and Utility of Neutrophil-to-Lymphocyte Ratio as a Predictive Biomarker for Advanced Non-Small-Cell Lung Cancer:A Prospective Observational Study.Clin Lung Cancer.2019;20(3):208-14 e2. Garon EB,Hellmann MD,Rizvi NA,Carcereny E,Leighl NB,Ahn MJ,et al.Five-Year Overall Survival for Patients With Advanced NonSmall-Cell Lung Cancer Treated With Pembrolizumab:Results From the Phase I KEYNOTE-001 Study.J Clin Oncol.2019;JCO1900934 Gupta D,Lis CG.Pretreatment serum albumin as a predictor of cancer survival:a systematic review of the epidemiological literature.Nutr J.2010;9:69. Kazandjian DG,Gong Y,Kazandjian H,Pazdur R,Blumenthal GM.Exploration of baseline derived neutrophil to lymphocyte ratio(dNLR)and lactate dehydrogenase(LDH)in patients(pts)with metastatic non-small cell lung cancer(mNSCLC)treated with immune checkpoint inhibitors(ICI)or cytotoxic chemotherapy(CCT).2018;36(15_suppl):3035. Khunger M,Patil PD,Khunger A,Li M,Hu B,Rakshit S,et al.Post-treatment changes in hematological parameters predict response to nivolumab monotherapy in non-small cell lung cancer patients.PLoS One.2018;13(10):e0197743. Luo M,Sun W,Wu C,Zhang L,Liu D,Li W,et al.High pretreatment serum gamma-glutamyl transpeptidase predicts an inferior outcome in nasopharyngeal carcinoma.Oncotarget.2017;8(40):67651-62. Mezquita L,Auclin E,Ferrara R,Charrier M,Remon J,et al.Association of the Lung Immune Prognostic Index With Immune Checkpoint Inhibitor Outcomes in Patients With Advanced Non-Small Cell Lung Cancer.JAMA Oncol.2018 Mar 1;4(3):351-357. Minami S,Ihara S,Ikuta S,Komuta K.Gustave Roussy Immune Score and Royal Marsden Hospital Prognostic Score Are Biomarkers of Immune-Checkpoint Inhibitor for Non-Small Cell Lung Cancer.World J Oncol.2019 Apr;10(2):90-100. Mok TSK,Wu YL,Kudaba I,Kowalski DM,Cho BC,Turna HZ,et al.Pembrolizumab versus chemotherapy for previously untreated,PD-L1-expressing,locally advanced or metastatic non-small-cell lung cancer(KEYNOTE-042):a randomised,open-label,controlled,phase 3 trial.Lancet.2019;393(10183):1819-30. Mulkey F,By K,Theoret MR,Maher VE,Pazdur R,Sridhara R.Analysis of early mortality in randomized clinical trials evaluating anti-PD-1 / PD-L1 antibodies:A systematic analysis by the United States Food and Drug Administration(FDA).2019;37(15_suppl):2516-. Petrelli F,Cabiddu M,Coinu A,Borgonovo K,Ghilardi M,Lonati V,et al.Prognostic role of lactate dehydrogenase in solid tumors:a systematic review and meta-analysis of 76 studies.Acta Oncol.2015;54(7):961-70. Powles T,Duran I,van der Heijden MS,Loriot Y,Vogelzang NJ,De Giorgi U et al.Atezolizumab versus chemotherapy in patients with platinum-treated locally advanced or metastatic urothelial carcinoma(IMvigor211):a multicentre,open-label,phase 3 randomised controlled trial.Lancet.2018;391(10122):748-57. Rizvi N,Cho BC,Reinmuth N,Lee KH,Ahn MJ,Luft A et al.Durvalumab with or without tremelimumab vs platinum-based chemotherapy as first-line treatment for metastatic non-small cell lung cancer:MYSTIC.ESMO Immuno-Oncology Congress;2018 December 13-16;Geneva,Switzerland.Abstract LBA6[oral presentation]. Sacdalan DB,Lucero JA,Sacdalan DL.Prognostic utility of baseline neutrophil-to-lymphocyte ratio in patients receiving immune checkpoint inhibitors:a review and meta-analysis.Onco Targets Ther.2018;11:955-65. Socinski MA,Jotte RM,Cappuzzo F,Orlandi F,Stroyakovskiy D,Nogami N,et al.Atezolizumab for First-Line Treatment of Metastatic Nonsquamous NSCLC.N Engl J Med.2018;378(24):2288-301. Shitara K,Ozguroglu M,Bang YJ,Di Bartolomeo M,Mandala M,Ryu MH,et al.Pembrolizumab versus paclitaxel for previously treated,advanced gastric or gastro-oesophageal junction cancer(KEYNOTE-061):a randomised,open-label,controlled,phase 3 trial.Lancet.2018;392(10142):123-33. Tham T,Bardash Y,Herman SW,Costantino PD.Neutrophil-to-lymphocyte ratio as a prognostic indicator in head and neck cancer:A systematic review and meta-analysis.Head Neck.2018;40(11):2546-57. Varga A,Bernard-Tessier A,Auclin E,Mezquita Perez L.Baldini C,et al.Applicability of the Lung Immune Prognostic Index(LIPI)in patients with metastatic solid tumors when treated with Immune Checkpoint Inhibitors(ICI)in early clinical trials.Ann Oncol.2019 Feb 1;30(Supplement 1). Wheler J,Tsimberidou AM,Hong D,Naing A,Falchook G,Piha-Paul S,et al.Survival of 1,181 patients in a phase I clinic:the MD Anderson Clinical Center for targeted therapy experience.Clin Cancer Res.2012;18(10):2922-9. Winquist E KS,Nichols AC,Lenehan JG,Macneil SD,Palma DA;.Early mortality with immune checkpoint inhibitors(IOs)in solid tumors:An inconvenient truth?J Clin Oncol.2018;36(suppl;abstr 12121).

[0246] Yu Y,Wang H,Yan A,Wang H,Li X,Liu J,et al.Pretreatment neutrophil to lymphocyte ratio in determining the prognosis of head and neck cancer:a meta-analysis.BMC Cancer.2018;18(1):383.

Claims

1. 1. A method of treating a patient having cancer, comprising: a) determining the concentrations of a plurality of clinicopathological markers in said patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; c) receiving a prediction from the gradient boosting machine learning model of whether the patient has a likelihood of dying within a 12-week period; and d) administering an anti-cancer therapeutic agent to the patient if the prediction indicates that the patient is unlikely to die within the 12 week period. The method includes:

2. The multiple clinicopathological markers were age, albumin (g / L), alanine phosphatase (U / L), alanine aminotransferase (U / L), aspartate aminotransferase (U / L), basophils (10 ３ / μL), basophils / white blood cells (%), bilirubin (μmol / L), body mass index, calcium (mmol / L), chloride (mmol / L), eosinophils (10 ３ / μL), eosinophils / leukocytes (%), gamma-glutamyltransferase (U / L), glucose (mmol / L), hematocrit (%), hemoglobin (g / L), potassium (mmol / L), lactate dehydrogenase (U / L), liver metastasis (yes / no), lymphocytes (10 ３ / μL), lymphocytes / white blood cells (%), magnesium (mmol / L), monocytes (10 ３ / μL), monocytes / leukocytes (%), neutrophils (10 ３ / μL), neutrophils / leukocytes (%), neutrophils / lymphocytes (%), platelets (10 ３ / μL), protein (g / L), gender, sodium (mmol / L), thyrotropin (mU / L), tumor size, uric acid (mmol / L) and white blood cells (10 ３ 13. The method of claim 1, comprising one or more of:

3. The multiple clinicopathological markers include neutrophils / white blood cells (%), neutrophils (10 ３ 3. The method of claim 2, comprising: IgG4 (g / μL), albumin (g / L), lactate dehydrogenase (U / L), aspartate aminotransferase (U / L) and / or gamma-glutamyltransferase (U / L).

4. The method of claim 1 , further comprising obtaining a sample from the patient.

5. 5. The method of claim 4, wherein the sample comprises a tissue biopsy, cerebrospinal fluid, lymphatic fluid, whole blood, serum, blood cells, urine, sweat, tears, saliva and / or feces.

6. The method of claim 5 , wherein the concentration of one or more of the plurality of clinical pathological markers is determined from the sample taken from the patient.

7. 2. The method of claim 1, wherein the anti-cancer therapeutic comprises cisplatin, gemcitabine, methotrexate, vinblastine, doxorubicin, cisplatin (MVAC), carboplatin, taxanes, temozolomide, dacarbazine, vinflunine, docetaxel, paclitaxel, nab-paclitaxel, vemurafenib, erlotinib, afatinib, cetuximab, bevacizumab, gefitinib, and / or pemetrexed.

8. 10. The method of claim 1, wherein the anti-cancer therapeutic comprises an immune checkpoint inhibitor.

9. The method of claim 8, wherein the immune checkpoint inhibitor comprises an anti-CTLA-4 antibody, an anti-PD-1 antibody, and / or an anti-PD-L1 antibody.

10. The method of claim 9, wherein the anti-CTLA-4 antibody is tremelimumab or ipilimumab.

11. 10. The method of claim 9, wherein the anti-PD-1 antibody is REGN2810, SHR1210, IBI308, PDR001, nivolumab, pembrolizumab, anti-PD-1, BGB-A317, BCD-100, or JS001.

12. The method of claim 9, wherein the anti-PD-L1 antibody comprises durvalumab, avelumab, atezolizumab, or KNO35.

13. 10. The method of claim 1, further comprising determining a cancer type of the patient and providing the determined cancer type to the gradient boosting machine learning model.

14. 1. A method of treating a patient having cancer, comprising: a) determining the concentrations of a plurality of clinicopathological markers in said patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; c) receiving a prediction from the gradient boosting machine learning model of whether the patient has poor immune fitness; and d) administering an anti-cancer therapeutic agent to the patient if the prediction indicates that the patient does not have reduced immune fitness. The method includes:

15. 15. The method of claim 14, further comprising: e) reducing at least one of the rate of cancer cell division, tumor growth, tumor size, tumor density, or rate of tumor metastasis in said patient.

16. 1. A method of treating a patient having cancer, comprising: a) determining the concentrations of a plurality of clinicopathological markers in said patient; b) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; c) receiving a prediction from the gradient boosting machine learning model of whether the patient has poor immune fitness; d) if said prediction indicates that said patient has reduced immune fitness, administering a therapeutic agent to said patient to enhance said patient's immune fitness; e) determining a second concentration of a plurality of clinicopathological markers for said patient; f) providing the determined second concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; g) receiving a second prediction from the gradient boosting machine learning model of whether the patient has poor immune fitness; and h) administering an anti-cancer therapeutic agent to the patient if the prediction indicates that the patient does not have reduced immune fitness. The method includes:

17. 1. A method for conducting a clinical trial of an anti-cancer therapeutic, comprising: a) identifying a potential participant for a clinical trial, the potential participant having cancer; b) determining the concentrations of a plurality of clinicopathological markers in said patient; c) providing the determined concentrations of the plurality of clinicopathological markers to a gradient boosting machine learning model; d) receiving a prediction from the gradient boosting machine learning model of whether the patient has a likelihood of dying within a 12-week period; and e) if said prediction indicates that said patient will not be likely to die within said 12 week period, including said patient in said clinical trial of said anti-cancer therapeutic agent. The method includes:

18. 18. The method of claim 17, wherein the anti-cancer therapeutic is an immune checkpoint inhibitor.

19. a computer memory configured to store model training data, the model training data including a first set of cancer patients undergoing cancer treatment, the model training data correlating (i) results from clinical tests performed on the first set of cancer patients and tumor types of the first set of cancer patients, with (ii) whether an individual patient from the first set of cancer patients died within a reference number of weeks from initiation of the cancer treatment; and one or more processors in communication with the computer memory, training a series of decision trees based on the model training data until a stopping condition is reached, each successive decision tree in the series being constructed based on the residual values ​​of its predecessor in the series; and creating a gradient boosting machine learning model as an additive function of the set of decision trees, the gradient boosting machine learning model predicting whether an additional cancer patient will die within the reference number of weeks based on the additive function applied to results from the clinical tests performed on the additional cancer patient and the tumor type of the additional cancer patient. one or more processors configured to execute program instructions to perform A computing system comprising:

20. 20. The computing system of claim 19, wherein the baseline number of weeks is 12 weeks.

21. 20. The computing system of claim 19, wherein the results from the laboratory test provide one or more of a neutrophil to lymphocyte ratio, a neutrophil concentration, an albumin concentration, a lactate dehydrogenase concentration, an aspartate aminotransferase concentration, or a gamma-glutamyltransferase concentration.

22. the neutrophil to lymphocyte ratio, the neutrophil concentration, the albumin concentration, the lactate dehydrogenase concentration, the aspartate aminotransferase concentration, and the gamma-glutamyltransferase concentration are selected from 20 or more clinical tests performed on the first set of cancer patients, said selecting comprising: training the gradient boosting machine learning model on the 20 or more clinical tests; For each respective clinical test, training each variation of the gradient boosting machine learning model on the 20 or more clinical tests, excluding the respective clinical test; and determining a predictive value for each of the twenty or more clinical tests by comparing the accuracy of the variations of the gradient boosting machine learning model; 22. The computing system of claim 21, comprising:

23. 20. The computing system of claim 19, wherein the tumor type is one of non-small cell lung cancer, urothelial carcinoma, squamous cell carcinoma of the head and neck, gastric cancer, renal cell carcinoma, or Hodgkin's lymphoma.

24. 20. The computing system of claim 19, wherein the additive function applies a multiplicative learning rate to an outcome from each of the sequence of decision trees.

25. 20. The computing system of claim 19, wherein predicting whether the additional cancer patient will die within the reference number of weeks comprises determining a probability that the additional cancer patient will die within the reference number of weeks by using the additive function.

26. 26. The computing system of claim 25, wherein predicting whether the additional cancer patient will die within the reference number of weeks further comprises placing the additional cancer patient in either a high mortality risk category or a low mortality risk category based on the probability and a cutoff value.

27. 27. The computing system of claim 26, wherein the cutoff value is selected such that there is no false positive rate greater than 5-20% when predicting whether the first set of cancer patients will die within the reference number of weeks.

28. 27. The computing system of claim 26, wherein cancer patients placed in the high mortality risk category are excluded from clinical trials for the treatment of cancer and cancer patients placed in the low mortality risk category are included in the clinical trials for the treatment of cancer.

29. 20. The computing system of claim 19, wherein the stopping condition is that the set of decision trees includes a reference number of decision trees.

30. 20. The computing system of claim 19, wherein the stopping condition is that the residual values ​​of a particular decision tree in the set of decision trees are all less than a reference value.

31. The computer memory is further configured to store model tuning data, the model tuning data including a second set of cancer patients receiving a cancer treatment, the model tuning data correlating (i) results from the clinical tests performed on the second set of cancer patients and tumor types of the second set of cancer patients with (ii) whether the second set of cancer patients died within the reference number of weeks from initiation of the cancer treatment, and the one or more processors: creating two or more variations of the gradient boosting machine learning model by training a set of decision trees for each variation having a different set of hyperparameters; applying the variations to the model training data and the model tuning data; and selecting, from the variations, as the gradient boosting machine learning model, a particular variation that provides within a reference accuracy a prediction of whether the first set of cancer patients and the second set of cancer patients will die within the reference number of weeks from initiation of the cancer treatment.

20. The computing system of claim 19, further configured to execute further program instructions for:

32. 32. The computing system of claim 31, wherein the accuracy of the benchmark is based on a true positive rate or a true negative rate of predicting whether the first set of cancer patients will die within the benchmark number of weeks from initiation of the cancer treatment.

33. 32. The computing system of claim 31 , wherein the accuracy of the benchmark is based on a p-value of an interaction between (i) a prediction of whether the second set of cancer patients will die within the benchmark number of weeks from initiation of the cancer treatment, and (ii) PD-L1 positive patients in the second set of cancer patients.

34. 32. The computing system of claim 31 , wherein the hyperparameters include one or more of: a maximum depth of the decision tree set to 2, 3, 4, 5, or 6; a minimum loss drop required to split a node in the decision tree set to 0, 1, 2, 3, or 4; a minimum sum of weights of all observations required at a child node in the decision tree set to between 3.0 and 7.0; a proportion of input parameters randomly sampled for each of the decision trees set to between 0.7 and 1.0; a proportion of observations randomly sampled for each of the decision trees set to between 0.4 and 0.8; and a number of decision trees in the set of decision trees set to between 10 and 25.

35. obtaining model training data comprising a first set of cancer patients undergoing cancer treatment, the model training data correlating (i) results from clinical tests performed on the first set of cancer patients with tumor types of the first set of cancer patients, and (ii) whether an individual patient from the first set of cancer patients died within a reference number of weeks from initiation of the cancer treatment; training a series of decision trees based on the model training data until a stopping condition is reached, each successive decision tree in the series being constructed based on the residual values ​​of its predecessor in the series; and creating a gradient boosting machine learning model as an additive function of the set of decision trees, the gradient boosting machine learning model predicting whether an additional cancer patient will die within the reference number of weeks based on an additive function applied to results from a laboratory test performed on the additional cancer patient and a tumor type of the additional cancer patient.

36. 36. The computer-implemented method of claim 35, wherein the baseline number of weeks is 12 weeks.

37. 36. The computer-implemented method of claim 35, wherein the results from the laboratory test provide one or more of a neutrophil to lymphocyte ratio, a neutrophil concentration, an albumin concentration, a lactate dehydrogenase concentration, an aspartate aminotransferase concentration, or a gamma-glutamyltransferase concentration.

38. the neutrophil to lymphocyte ratio, the neutrophil concentration, the albumin concentration, the lactate dehydrogenase concentration, the aspartate aminotransferase concentration, and the gamma-glutamyltransferase concentration are selected from 20 or more clinical tests performed on the first set of cancer patients, said selecting comprising: training the gradient boosting machine learning model on the 20 or more clinical tests; For each respective clinical test, training each variation of the gradient boosting machine learning model on the 20 or more clinical tests, excluding the respective clinical test; and determining a predictive value for each of the twenty or more clinical tests by comparing the accuracy of the variations of the gradient boosting machine learning model; 38. The computer-implemented method of claim 37, comprising:

39. 36. The computer-implemented method of claim 35, wherein the tumor type is one of non-small cell lung cancer, urothelial carcinoma, squamous cell carcinoma of the head and neck, gastric cancer, renal cell carcinoma, or Hodgkin's lymphoma.

40. 36. The computer-implemented method of claim 35, wherein the additive function applies a multiplicative learning rate to an outcome from each of the sequence of decision trees.

41. 36. The computer-implemented method of claim 35, wherein predicting whether the additional cancer patient will die within the reference number of weeks comprises determining a probability that the additional cancer patient will die within the reference number of weeks by using the additive function.

42. 42. The computer-implemented method of claim 41, wherein predicting whether the additional cancer patient will die within the reference number of weeks further comprises placing the additional cancer patient in either a high mortality risk category or a low mortality risk category based on the probability and a cutoff value.

43. 43. The computer-implemented method of claim 42, wherein the cutoff value is selected such that there is no false positive rate greater than 5-20% when predicting whether the first set of cancer patients will die within the reference number of weeks.

44. 43. The computer-implemented method of claim 42, wherein cancer patients who are placed in the high mortality risk category are excluded from clinical trials for treatments for cancer and cancer patients who are placed in the low mortality risk category are included in the clinical trials for treatments for cancer.

45. 36. The computer-implemented method of claim 35, wherein the stopping condition is that the set of decision trees includes a reference number of decision trees.

46. 36. The computer-implemented method of claim 35, wherein the stopping condition is that the residual values ​​of a particular decision tree in the set of decision trees are all below a reference value.

47. obtaining model tuning data, the model tuning data including a second set of cancer patients receiving a cancer treatment, the model tuning data (i) correlating results from the clinical tests performed on the second set of cancer patients with tumor types of the second set of cancer patients, and (ii) whether the second set of cancer patients died within the reference number of weeks of initiation of the cancer treatment; creating two or more variations of the gradient boosting machine learning model by training a set of decision trees for each variation having a different set of hyperparameters; applying the variations to the model training data and the model tuning data; and selecting, from the variations, as the gradient boosting machine learning model, a particular variation that provides within a reference accuracy a prediction of whether the first set of cancer patients and the second set of cancer patients will die within the reference number of weeks from initiation of the cancer treatment.

36. The computer-implemented method of claim 35, further comprising:

48. 48. The computer-implemented method of claim 47, wherein the accuracy of the criteria is based on a true positive rate or a true negative rate of predicting whether the first set of cancer patients will die within the criteria number of weeks from initiation of the cancer treatment.

49. 48. The computer-implemented method of claim 47, wherein the accuracy of the criteria is based on a p-value of the interaction between (i) a prediction of whether the second set of cancer patients will die within the criteria number of weeks from initiation of the cancer treatment, and (ii) PD-L1 positive patients in the second set of cancer patients.

50. 48. The computer-implemented method of claim 47, wherein the hyper-parameters include one or more of: a maximum depth of the decision tree set to 2, 3, 4, 5, or 6; a minimum loss drop required to split a node in the decision tree set to 0, 1, 2, 3, or 4; a minimum sum of weights of all observations required at a child node in the decision tree set from 3.0 to 7.0; a proportion of input parameters randomly sampled for each of the decision trees set from 0.7 to 1.0; a proportion of observations randomly sampled for each of the decision trees set from 0.4 to 0.8; and a number of decision trees in the set of decision trees set from 10 to 25.

51. 51. An article of manufacture comprising a non-transitory computer readable medium having stored thereon program instructions which, when executed by a computing device, cause the computing device to perform the operations of any one of claims 35 to 50.

52. a computer memory configured to store a gradient boosting machine learning model, the gradient boosting machine learning model trained with model training data, the model training data including a first set of cancer patients undergoing cancer treatment, the model training data relating (i) results from clinical tests performed on the first set of cancer patients and tumor types of the first set of cancer patients with (ii) whether an individual patient from the first set of cancer patients died within a reference number of weeks from initiation of the cancer treatment, wherein training the gradient boosting machine learning model included (i) training a series of decision trees until a stopping condition is reached, each successive decision tree in the series constructed based on residual values ​​of its predecessor in the series, and (ii) determining an additive function for the series of decision trees; and one or more processors in communication with the computer memory, obtaining results from said clinical test performed on additional cancer patients; obtaining a tumor type of said further cancer patient; applying the gradient boosting machine learning model to results from the clinical test performed on the additional cancer patient and to the tumor type of the additional cancer patient; and receiving a prediction from the gradient boosting machine learning model whether the additional cancer patient will die within the reference number of weeks. one or more processors configured to execute program instructions to perform A computing system comprising:

53. 53. The computing system of claim 52, wherein the baseline number of weeks is 12 weeks.

54. 53. The computing system of claim 52, wherein the results from the laboratory test provide one or more of a neutrophil to lymphocyte ratio, a neutrophil concentration, an albumin concentration, a lactate dehydrogenase concentration, an aspartate aminotransferase concentration, or a gamma-glutamyltransferase concentration.

55. 53. The computing system of claim 52, wherein the tumor type is one of non-small cell lung cancer, urothelial carcinoma, squamous cell carcinoma of the head and neck, gastric cancer, renal cell carcinoma, or Hodgkin's lymphoma.

56. 53. The computing system of claim 52, wherein the additive function applies a multiplicative learning rate to an outcome from each of the sequence of decision trees.

57. 53. The computing system of claim 52, wherein predicting whether the additional cancer patient will die within the reference number of weeks comprises determining a probability that the additional cancer patient will die within the reference number of weeks by using the additive function.

58. 58. The computing system of claim 57, wherein predicting whether the additional cancer patient will die within the reference number of weeks further comprises placing the additional cancer patient in either a high mortality risk category or a low mortality risk category based on the probability and a cutoff value.

59. 59. The computing system of claim 58, wherein the cutoff value is selected such that there is no false positive rate greater than 5-20% when predicting whether the first set of cancer patients will die within the reference number of weeks.

60. 59. The computing system of claim 58, wherein cancer patients placed in the high mortality risk category are excluded from clinical trials for the treatment of cancer and cancer patients placed in the low mortality risk category are included in the clinical trials for the treatment of cancer.

61. 53. The computing system of claim 52, wherein the stopping condition is that the set of decision trees includes a reference number of decision trees.

62. 53. The computing system of claim 52, wherein the stopping condition is that the residual values ​​of a particular decision tree in the set of decision trees are all less than a reference value.

63. the gradient boosting machine learning model has also been trained using model tuning data, the model tuning data including a second set of cancer patients undergoing cancer treatment, the model tuning data (i) associating results from the clinical tests performed on the second set of cancer patients with tumor types of the second set of cancer patients with (ii) whether the second set of cancer patients died within the reference number of weeks of initiation of the cancer treatment; training the gradient boosting machine learning model using the model tuning data comprises: creating two or more variations of the gradient boosting machine learning model by training a set of decision trees for each variation having a different set of hyperparameters; applying the variations to the model training data and the model tuning data; and selecting, from the variations, as the gradient boosting machine learning model, a particular variation that provides within a reference accuracy a prediction of whether the first set of cancer patients and the second set of cancer patients will die within the reference number of weeks from initiation of the cancer treatment.

53. The computing system of claim 52, comprising:

64. 64. The computing system of claim 63, wherein the accuracy of the standard is based on a true positive rate or a true negative rate of predicting whether the first set of cancer patients will die within the standard number of weeks from initiation of the cancer treatment.

65. 64. The computing system of claim 63, wherein the accuracy of the reference is based on a p-value of an interaction between (i) a prediction of whether the second set of cancer patients will die within the reference number of weeks from initiation of the cancer treatment, and (ii) PD-L1 positive patients in the second set of cancer patients.

66. 64. The computing system of claim 63, wherein the hyperparameters include one or more of: a maximum depth of the decision tree set to 2, 3, 4, 5, or 6; a minimum loss drop required to split a node in the decision tree set to 0, 1, 2, 3, or 4; a minimum sum of weights of all observations required at a child node in the decision tree set to between 3.0 and 7.0; a proportion of input parameters randomly sampled for each of the decision trees set to between 0.7 and 1.0; a proportion of observations randomly sampled for each of the decision trees set to between 0.4 and 0.8; and a number of decision trees in the set of decision trees set to between 10 and 25.

67. obtaining a gradient boosting machine learning model, the gradient boosting machine learning model trained with model training data, the model training data including a first set of cancer patients undergoing cancer treatment, the model training data relating (i) results from clinical tests performed on the first set of cancer patients and tumor types of the first set of cancer patients with (ii) whether each patient from the first set of cancer patients died within a reference number of weeks from initiation of the cancer treatment, training the gradient boosting machine learning model including (i) training a series of decision trees until a stopping condition is reached, each successive decision tree in the series constructed based on residual values ​​of its predecessor in the series, and (ii) determining an additive function of the series of decision trees; obtaining results from said clinical test performed on additional cancer patients; obtaining a tumor type of said further cancer patient; applying the gradient boosting machine learning model to results from the clinical test performed on a further cancer patient and to the tumor type of the further cancer patient; and receiving a prediction from the gradient boosting machine learning model whether the additional cancer patient will die within the reference number of weeks.

23. A computer-implemented method comprising:

68. 68. The computer-implemented method of claim 67, wherein the baseline number of weeks is 12 weeks.

69. 68. The computer-implemented method of claim 67, wherein the results from the laboratory test provide one or more of a neutrophil to lymphocyte ratio, a neutrophil concentration, an albumin concentration, a lactate dehydrogenase concentration, an aspartate aminotransferase concentration, or a gamma-glutamyltransferase concentration.

70. 68. The computer-implemented method of claim 67, wherein the tumor type is one of non-small cell lung cancer, urothelial carcinoma, squamous cell carcinoma of the head and neck, gastric cancer, renal cell carcinoma, or Hodgkin's lymphoma.

71. 68. The computer-implemented method of claim 67, wherein the additive function applies a multiplicative learning rate to an outcome from each of the sequence of decision trees.

72. 68. The computer-implemented method of claim 67, wherein predicting whether the additional cancer patient will die within the reference number of weeks comprises determining a probability that the additional cancer patient will die within the reference number of weeks by using the additive function.

73. 73. The computer-implemented method of claim 72, wherein predicting whether the additional cancer patient will die within the reference number of weeks further comprises placing the additional cancer patient in either a high mortality risk category or a low mortality risk category based on the probability and a cutoff value.

74. 74. The computer-implemented method of claim 73, wherein the cutoff value is selected such that there is no false positive rate greater than 5-20% when predicting whether the first set of cancer patients will die within the reference number of weeks.

75. 74. The computer-implemented method of claim 73, wherein cancer patients who are placed in the high mortality risk category are excluded from clinical trials for treatment of cancer and cancer patients who are placed in the low mortality risk category are included in the clinical trials for treatment of cancer.

76. 68. The computer-implemented method of claim 67, wherein the stopping condition is that the set of decision trees includes a reference number of decision trees.

77. 68. The computer-implemented method of claim 67, wherein the stopping condition is that the residual values ​​of a particular decision tree in the set of decision trees are all below a reference value.

78. The gradient boosting machine learning model has also been trained using model tuning data, the model tuning data including a second set of cancer patients undergoing cancer treatment, the model tuning data associating (i) results from the clinical tests performed on the second set of cancer patients and tumor types of the second set of cancer patients with (ii) whether the second set of cancer patients died within the reference number of weeks from initiation of the cancer treatment, and training the gradient boosting machine learning model using the model tuning data comprises: creating two or more variations of the gradient boosting machine learning model by training a set of decision trees for each variation having a different set of hyperparameters; applying the variations to the model training data and the model tuning data; and selecting, from the variations, as the gradient boosting machine learning model, a particular variation that provides within a reference accuracy a prediction of whether the first set of cancer patients and the second set of cancer patients will die within the reference number of weeks from initiation of the cancer treatment.

68. The computer-implemented method of claim 67, comprising:

79. 79. The computer-implemented method of claim 78, wherein the accuracy of the criteria is based on a true positive rate or a true negative rate of predicting whether the first set of cancer patients will die within the criteria number of weeks from initiation of the cancer treatment.

80. 79. The computer-implemented method of claim 78, wherein the accuracy of the criteria is based on a p-value of the interaction between (i) a prediction of whether the second set of cancer patients will die within the criteria number of weeks from initiation of the cancer treatment, and (ii) PD-L1 positive patients in the second set of cancer patients.

81. 79. The computer-implemented method of claim 78, wherein the hyper-parameters include one or more of: a maximum depth of the decision tree set to 2, 3, 4, 5, or 6; a minimum loss reduction required to split a node in the decision tree set to 0, 1, 2, 3, or 4; a minimum sum of weights of all observations required at a child node in the decision tree set from 3.0 to 7.0; a proportion of input parameters randomly sampled for each of the decision trees set from 0.7 to 1.0; a proportion of observations randomly sampled for each of the decision trees set from 0.4 to 0.8; and a number of decision trees in the set of decision trees set from 10 to 25.

82. 82. An article of manufacture comprising a non-transitory computer readable medium having stored thereon program instructions which, when executed by a computing device, cause the computing device to perform the operations of any one of claims 67 to 81.

Images