Integrating multiple data sources to predict all-cause readmission or mortality in patients with substance misuse

Area of Science:

• Health Informatics
• Data Science in Healthcare
• Public Health

Background:

• Patients with substance misuse admitted to hospitals have increased risks of adverse outcomes, including readmission and mortality.
• Early identification of high-risk individuals is crucial for timely interventions and improved healthcare resource optimization.

Purpose of the Study:

• To develop and evaluate methods for identifying patients with substance misuse at high risk of 30-day death or hospital readmission post-discharge.
• To explore the integration of diverse data sources for enhanced risk prediction.

Main Methods:

• Leveraged the Substance Misuse Data Commons for predictive modeling.
• Compared various machine learning algorithms using structured electronic health record (EHR) data, unstructured clinical notes, socioeconomic data, and emergency medical services (EMS) data.
• Developed a gradient-boosted machine model integrating structured EHR, socioeconomic, and EMS data.

Main Results:

• The gradient-boosted machine model combining structured EHR, socioeconomic, and EMS data achieved the best performance (c-statistic 0.746).
• This integrated model outperformed other machine learning approaches and data source combinations.
• Incorporating unstructured clinical notes did not significantly improve predictive performance.

Conclusions:

• Multi-source data integration, extending beyond traditional EHRs, can significantly enhance risk assessment for hospitalized patients with substance misuse.
• Prior hospitalizations, EMS encounters, and discharge disposition were key predictors of adverse outcomes.
• Further research is needed to effectively utilize unstructured clinical data for risk prediction.