Machine-learning (ML) model predictive for chemotherapy toxicity and termination in a real-world community oncology practice.

Abstract

e13697 Background: Randomized clinical trials involve carefully selected patient (pts) to develop new therapeutic indications. However, effectiveness in real-world community oncology setting is often limited as many pts are unable to undergo standard-of-care treatments. We utilized machine learning in multiple real-world community oncology practices to develop a risk model for predicting treatment discontinuation due to chemo-related toxicities. Methods: Several large language models (LLMs) were used to extract oncology-related clinical features and outcomes from de-identified Electronic Medical Records (EMRs). LLMs included Mistral and LLaMA 3.2. The extraction efficiency of each model was evaluated for clinical features that included pt. demographics, drugs and toxicities amongst other 125 other clinical features. Pts initiating standard-of-care chemotherapy for four metastatic tumor types, including metastatic castration-resistant prostate cancer (mCRPC), metastatic colon cancer (mCC), metastatic lung cancer (mLC), and metastatic breast cancer (mBC) were categorized based on age (< / > 70 years). Clinical outcomes (labels) analyzed based in the two age categories included incidence of chemotherapy discontinuation due to drug toxicity. A predictive model based on pre-treatment clinical features was built using Cox Proportional Hazard Regression (CPHR) models for each tumor type and features with univariate Hazard Ratios (HRs significant at P < 0.05) were included in an aggregate risk model that determined discontinuation of chemotherapy. Model performance was assessed by determining Area Under the Curve (AUC). Institutional Review Board (IRB) approval was obtained prior to research. Results: 3,537 metastatic cancer pts initiating chemotherapies were identified from several community oncology practices, which included mCRPC (N = 1033), mCC (1176), mLC (718), and mBC (610) pts. Among all LLMs, LlaMa3 was the best for extracting clinical features from unstructured EMR data with a F1 score of 86% and reasons for treatment discontinuation with an accuracy of 84%. Table 1 shows for each tumor type, the number of pts < / > 70 years whose chemotherapy treatments were discontinued due to toxicity, along with a median time-to-treatment discontinuation. A risk model was developed for metastatic pts which highlighted drug induced toxicity related chemotherapy discontinuation within the first 30 days from the start of the treatment. The top four clinical features leading to discontinuation were pneumonia (HR: 5.6), peripheral vascular disease (HR:3.1), anemia (HR:2.36), acute kidney failure (HR: 2.25) and the AUC of the aggregated model for predicting discontinuation of chemotherapy included into a nomogram was observed at 0.67. Conclusions: In this training set, LLMs demonstrated a high degree of efficiency and accuracy in extracting clinical features that can be used for risk prediction for drug toxicity and discontinuation in elderly patients, especially in mCRPC pts. Age distribution of metastatic cancer patients and chemotherapy discontinuation rates. mCC mLC mBC mCRPC Patients above 70 years 246 577 247 664 Discontinuation due to toxicity 165 360 233 350 Discontinuation due to toxicity within 30 days 24 (15%) 62 (17%) 6 (2.6%) 42 (12%) Median time to discontinuation (days) 84 76 223 182 Patients under 70 years 472 599 363 369 Discontinuation due to toxicity 343 408 140 194 Median time to discontinuation (days) 96 92 175 136