Predicting Optimal Treatment Regimen to Improve Outcomes of Patients With CLL/SLL Using Random Survival Forest

Speaker(s)

Butkowski D1, Cui Z2, Khanal M2, Lipkovich I2, Kadziola Z2, Faries DE2, Choong C2, Chen Y2, Bhandari NR2, Hess L2
1Florida International University, North Miami, FL, USA, 2Eli Lilly and Company, Indianapolis, IN, USA

Presentation Documents

OBJECTIVES: To predict optimal treatment in first and second line of therapy (LOT) that maximizes overall survival (OS) in patients with chronic lymphocytic leukemia (CLL) or small lymphocytic lymphoma (SLL) by using random survival forest (RSF) for clinical decision-making.

METHODS: This study used the nationwide Flatiron Health electronic health record (EHR)-derived de-identified database. Eligible patients were adults diagnosed with CLL/SLL who received ≥1 LOT between 01JAN2016 and 31AUG2023. Individual regimens were grouped into hierarchy regimen classes; the five most common were included in this analysis. Study cohorts were randomly partitioned 1000 times into 80% training and 20% validation subsets. RSF models were used to predict optimal regimen classes in first and second LOT based on baseline demographics and clinical characteristics. The OS expected under the predicted treatment regimen was compared to that under the current prescribing practice by using Cox proportional hazards regression, adjusted for baseline characteristics imbalance by inverse probability weighting.

RESULTS: The study cohort included 7,219 and 2,252 patients with first and second LOTs, respectively. RSF models suggested that 20.5% of patients received optimal treatment in both LOTs. To maximize OS, RSF model recommended greater use of BCL2i + antiCD20, increasing from observed 6% of patients to estimated optimal 35% in first LOT and from observed 17% of patients to estimated optimal 74% in second LOT. The estimated gains in OS from using the optimal treatment, in terms of restricted mean survival time, were 0.2 and 0.6 years, with hazard ratios (95% prediction interval) 0.86 (0.67, 1.1) and 0.76 (0.42, 1.2) in first and second LOT, respectively.

CONCLUSIONS: RSF was feasible using oncology EHR data, building the evidence to inform how machine learning may provide recommendations for oncologists in choosing treatments that improve outcomes for patients with CLL/SLL.

Code

MSR83

Topic

Methodological & Statistical Research

Topic Subcategory

Artificial Intelligence, Machine Learning, Predictive Analytics, Confounding, Selection Bias Correction, Causal Inference

Disease

Drugs, Oncology