Development and temporal validation of an interpretable point score for in-hospital mortality in CLL/SLL using the...

In-hospital mortality among patients with chronic lymphocytic leukemia/small lymphocytic lymphoma (CLL/SLL) is ∼6%, yet no validated, transparent bedside tool exists to guide acute care. We sought to develop and benchmark interpretable mortality prediction models using routine administrative data. We analyzed 117,765 CLL/SLL admissions in the 2016-2022 U.S. National Inpatient Sample. Fifty-nine ICD-10 predictors - 52 chronic Elixhauser/Charlson comorbidities and 7 acute complications - were entered into survey-weighted ridge logistic regression (Model A), grouped LASSO with integer scaling (Model B), a non-negative truncation of Model B (Model C), and eight machine learning comparators. Temporal validation used 2021-2022 data. We assessed discrimination, calibration, decision-curve net benefit, and performance across 23 subgroups. Crude mortality was 5.8% (6781/117,765). Model A achieved an AUROC of 0.851 with excellent calibration, while the best machine learning model improved AUROC trivially. Model B (42 items) and Model C (23 items, 0-116 points) retained AUROCs of 0.846 and 0.842, respectively; mortality across Model C quartiles ranged from < 1% to ∼22%. Across all 23 subgroups, AUROCs remained ≥ 0.75 with minimal inter-cccccmodel variation. Decision curves showed net benefit over "treat-all" and "treat-none" strategies for predicted risks between 5% and 25%; at a 10% threshold, Model B flagged 90 per 1000 admissions and captured 57% of deaths. Regularized logistic regression yields an equitable, well-calibrated CLL/SLL mortality score that matches complex machine learning models while maximizing interpretability. The 23-item non-negative score is readily deployable for triage; marginal ML gains may not justify added complexity.