Background
A large Medicaid health plan serving over 1.5 million members identified a critical need to improve care coordination, reduce costs, and enhance patient outcomes. The plan recognized that accurately predicting the Length of Stay (LoS) for hospital admissions could be a key factor in achieving these goals.
Challenge
The health plan faced several obstacles in implementing an effective LoS prediction system:
- Diverse patient population with complex health needs
- Limited access to real-time clinical data and pro-active risk stratification at and during the inpatient admission
- Variability in provider practices and hospital systems within the network
- Lack of standardized data collection and reporting across different care settings
- Resource constraints in terms of skilled personnel to analyze large volumes of healthcare data
The goal was to develop a predictive model that could identify patients at risk of extended hospital stays, enable proactive care management, optimize resource allocation, reduce unnecessary hospital days, and improve overall patient outcomes.
Solution
The health plan partnered with a healthcare analytics firm to develop a machine learning model for LoS prediction. The approach included:
- Data Integration: Collected data from claims, authorization systems, HL7 ADTs, population level SDoH data, risk scores (ACG, HPI), geolocation and weather APIs to prepare a dataset ready for ML.
- Model Development: Developed a Weighted XGBoost algorithm, optimized for high recall to minimize false negatives.
- Feature Selection: Identified key predictors such as demographics, diagnoses, procedures, comorbidities, prior utilization, and social determinants of health, transportation barriers, hospital operational characteristics, environmental factors like weather at the time of admission.
- Risk Stratification: Categorized patients into low, medium, and high risk for extended stays.
- Integration with Care Management Systems: Implemented the model to provide LoS estimates at the time of admission and the data is served on a PowerBI dashboard for care teams and administrative staff.
Implementation
The predictive model was integrated into the health plan’s care management system, providing LoS predictions and risk assessments for members requiring hospital admission. Care managers used this information to:
- Prioritize high-risk cases for intensive care coordination
- Collaborate with hospitals on discharge planning from day one of admission
- Allocate resources more effectively across the member population
Results:
The LoS prediction model delivers measurable improvements across operational, financial, and clinical performance indicators.
Conclusion:
By implementing a predictive LoS model, the Medicaid health plan significantly improved its ability to manage member care, allocate resources efficiently, and enhance overall health outcomes. The early identification of members at risk for extended stays allowed for proactive interventions, resulting in better health outcomes and substantial cost savings.
Next Steps:
To further enhance the model’s effectiveness and achieve more comprehensive member care, the health plan should consider:
- Integrating HL7 data to improve interoperability and data exchange with hospital systems.
- Incorporating CCDA information to access standardized summaries of patient information, ensuring a more accurate and complete medical history for each member.
- Exploring advanced AI techniques to refine predictions and agentic automation certain aspects of care coordination.
- Enhancing data sharing agreements with network providers to access more real-time clinical data.
- Continuously monitoring and refining the model to adapt to changing population health trends and healthcare delivery patterns.
These next steps will help the health plan move towards a more holistic view of its members, potentially uncovering further improvements in care quality and cost management.
