Risk Stratification Index 3.0, a Broad Set of Models for Predicting Adverse Events during and after Hospital...

Area of Science:

• Health Informatics
• Clinical Prediction Modeling
• Healthcare Management

Background:

• Risk stratification is crucial for guiding clinical decisions and optimizing patient care.
• Predictive tools are needed to identify hospitalized patients at risk for adverse events and high healthcare utilization.
• Administrative claims data, including International Classification of Diseases, Tenth Revision (ICD-10) codes, offer a rich source for developing such tools.

Purpose of the Study:

• To develop and validate a suite of predictive models using ICD-10 diagnostic and procedural codes.
• To predict adverse events and care utilization outcomes for hospitalized patients.
• To assess the performance of these models in stratifying patient risk.

Main Methods:

• Models were developed using Medicare admissions data from 2017-2018, incorporating patient demographics and historical coding data.
• Predictive features were used to forecast outcomes through 90 days post-admission, including unplanned admissions, mortality, and major complications.
• Model performance was evaluated on 2019 out-of-sample data, comparing logistic regression with machine learning methods.

Main Results:

• The models demonstrated good predictive performance on the validation set, with an average area under the curve of 0.76.
• Strong model calibration was observed, particularly for identifying patients at lowest risk (R² of 1.00).
• Predictive accuracies between logistic regression and machine learning techniques were generally similar, suggesting a ceiling for claims data prediction.

Conclusions:

• Predictive analytical modeling using administrative claims history can generate individualized risk profiles at hospital admission.
• These risk profiles can aid in guiding patient management and potentially improve clinical care pathways.
• The study identified the value and limitations of predictive information derived from medical claims data.