(1343) Predicting Chemotherapy Eligibility for Patients with Colorectal Cancer: Using Machine Learning Methods Based on Claims Data Linked to Electronic Medical Records

Objectives: Colorectal cancer (CRC) is the second leading cause of cancer-related deaths worldwide. In 2020, it was estimated that more than 1.9 million new cases of CRC arose. Although surgery is the backbone of curative treatment in CRC, surgery is not appropriate in most cases and adjuvant chemotherapy is recommended to prolong survival and reduce symptoms. However, significant variations exist in the effectiveness of chemotherapy use, and treatment-related side effects are a significant concern because not all patients can tolerate chemotherapies, suggesting the importance of personalized medicine for chemotherapy. As a subsection of artificial intelligence, machine learning (ML) is increasingly used in clinical oncology to inform treatment planning based on real-world big data. The objective of this study was to 1) develop an algorithm to accurately identify patients that are eligible for chemotherapy among patients with colorectal cancer using machine learning methods, and 2) evaluate the performance of the developed algorithms using different machine learning models.

Methods: This study employed a retrospective cohort study design to predict chemotherapy among patients with colorectal cancer (CRC). Participants were obtained using a large claims data linked to the Electronic Medical Records between 2012 and 2020, representing a comprehensive and longitudinal repository of healthcare records of more than millions of enrollees. T-tests and chi-squared tests were used to compare baseline characteristics between CRC patients with or without chemotherapy. Patients who met the eligibility criteria needed to have either an electronic medical record (EMR) encounter related to CRC or a diagnostic code for CRC from the International Classification of Diseases, Ninth Revision and Tenth Revision, Clinical Modification (ICD-9-CM/ICD-10-CM). To construct our predictive models, we used six machine learning algorithms, including logistic regression, random forest (RF), gradient boosting machine (GBM), LightGBM, and eXtreme Gradient Boosting (XGBoost). To construct validated models, we divided the study population into two distinct subsets, including a training set and a test set. To prevent any overlap, the division was randomized at the patient level. The training dataset comprised 66.7% of the total population, while the validation dataset comprised 33.3%. After training, we evaluated the algorithm's performance using the test dataset. Performance metrics included the area under the receiver operating characteristic curve (AUROC), accuracy (ACC), sensitivity, and specificity.

Results: The top five predictors for chemotherapy eligibility listed in order were radiation (1.18), surgery (0.67), mild liver disease (0.60), metastasis (0.52), and age (0.5). For model performance based on the AUROC, LightGBM emerges as the top-performing algorithm with an impressive score of 0.826. It is closely followed by Random Forest, which secures the second position with an AUC of 0.825, indicating strong discriminatory capabilities. In terms of accuracy, LightGBM also leads the models (0.781), closely followed by Random Forest (0.777) and GBM (0.772. In terms of sensitivity, XGBoost stands out as the top performer (0.703).

Conclusions: In the era of digital health and big data, this study makes an important contribution to patient care. Predicting chemotherapy eligibility is a turning point for future treatment regimens because of great variations in the effectiveness and tolerance of chemotherapies. It is reported that machine learning methods can improve therapy accuracy 15% to 25%. It can also help improve our basic understanding of chemotherapies.

Using ML technologies, health care providers (e.g., pharmacists, clinicians) can make informed and evidence-based decisions for the best regimen plans possible, which can lead to more favorable patient outcomes as well as cost savings. This study also shows the importance of Pharmacists as Personalized Medicine Experts (PRIME) in daily practice. Future studies with similar strategies will be needed for patients with other cancers.