(285) Risk Factor Identification Using Machine Learning and High-Dimensional Hospital Data to Inform Propensity and Disease Risk Score Models: An Applied Example in Hemostasis-Related Surgical Complications

Publication Number: 959

Background: Propensity score (PS) and disease risk score (DRS) methods are widely used to reduce bias in causal inference based on retrospective analyses of healthcare data. However, these methods hinge on the inclusion of relevant confounders/risk factors, for which there may be only minimal a priori knowledge.

Objectives: To describe an approach to identify risk factors in high-dimensional hospital data to inform PS/DRS models. Taking an approach in which we favor clinical face validity of risk factors over “black box” predictive analytic approaches, we search for risk factors for hemostasis-related surgical complications (HRSC), a difficult-to-predict outcome for which there is limited prior literature on risk factors.

Methods: Retrospective study of the Premier Healthcare Database. Patients included for study underwent video-assisted thoracoscopic lobectomy (VATL), laparoscopic right colectomy (LRC), or laparoscopic sleeve gastrectomy (LSG) on an inpatient setting between Oct-2015 and Feb-2020 (first=index). For each cohort, a high-dimensional dataset of candidate risk factors was built based on all observed secondary diagnoses recorded during index and designated as “present on admission”; as this resulted in >5,000 unique diagnoses, the Clinical Classification Software Refined (CCSR) was used to aggregate diagnoses into >500 clinically-meaningful categories. The outcome, HRSC, comprised hemorrhage, control of bleeding, and acute posthemorrhagic anemia. Candidate risk factors were fed into regularized logistic regression models with a 70%/30% train/test split for each cohort; clinically-plausible risk factors that were consistently significant predictors of HRSC across the 3 training models were then used in a final parsimonious model including sex, age, race, and payor; finally, the parsimonious model was applied to the test data to compare predicted risk with observed incidence of HRSC.

Results: The study included 11,141 VATL, 20,156 LRC, and 121,547 LSG patients, in whom 7.5%, 7.8%, and 1.2% experienced HRSC, respectively. Ultimately, 9 clinically-plausible CCSR categories were identified as being statistically significant predictors across all 3 cohorts (e.g., coagulation and hemorrhagic disorders, malnutrition, alcohol-related disorders, among others) In the parsimonious model applied to the test data, the observed incidence of HRSC was substantially higher in the top quintile vs. bottom quintile of predicted risk: LSG 2.05% vs. 0.53%, LRC 13.30% vs. 4.11%, VATS 12.49% vs. 5.04%.

Conclusions: Machine learning methods can be applied to high-dimensional data to identify individual risk factors for outcomes that may not have been obvious a priori candidates.