For clinical operations leaders evaluating predictive analytics in post-acute care.
Most descriptions of predictive models in healthcare stay at a level of abstraction that does not help clinical leaders evaluate what they are buying. The following examples show what a purpose-built predictive model for post-acute care looks like from the inside — the inputs, the training data, the output structure, and the validation layer.
Mosai Transitions: A Home Health Mortality Risk Model
Mosai Transitions is a predictive model built specifically for home health clinical management. Its primary function is to predict the likelihood of patient mortality within approximately ninety days, using longitudinal clinical data collected across a patient's episodes of care.
What the Model Was Trained On
The Transitions model was trained on more than 460,000 patient records drawn exclusively from home health clinical populations. Post-acute-specific training matters because the clinical variables that predict outcomes in home health are structurally different from those in acute or ambulatory care.
What the Model Evaluates
For each active patient, the model evaluates hundreds of clinical variables: OASIS assessments at start of care, recertification, and discharge; skilled nursing visit documentation across every encounter; vital sign trends evaluated across time; functional and cognitive decline patterns; disease burden and progression; and clinician narrative notes processed through natural language processing to extract signals from free-text documentation
How the Output Works
The model generates a continuously refreshed risk estimate for each patient. As new clinical information is documented, the model recalculates. The output is not a static score assigned at episode start — it is a living risk estimate that reflects the patient's current trajectory. At the census level, a clinical supervisor sees a risk-ranked view of their active patient population that reflects the most recent available clinical data for each patient.
Mosai Muse: Predictive Models for Hospice
Mosai Muse applies predictive modeling to hospice clinical management across two distinct use cases, each requiring its own training data and model architecture.
Transition Risk Model
The Muse Transition Risk Model was trained on more than 4.5 million hospice patient visits. For each patient, it analyzes approximately 900 patient-specific data points: 93 structured numerical clinical features, NLP-processed narrative documentation, and approximately 200 derived text-based features. The model predicts mortality likelihood within ten days of the most recent skilled visit and assigns each patient to one of eight risk categories — giving supervisors a more precise basis for allocating care intensity across a hospice caseload than a binary high/low classification.
Revocation Risk Model
The Muse Revocation Risk Model addresses a distinct problem: identifying patients at elevated likelihood of disenrolling from hospice services before death. Trained on more than 246,000 hospice patient records and analyzing approximately 281 data points per patient, the model identifies patients approximately five times more likely to revoke hospice services within thirty days — giving care teams an actionable window to intervene before a revocation occurs.
What Both Models Have in Common
Despite serving different care settings and predicting different outcomes, both models share the structural commitments that distinguish them from general-population AI applied to post-acute care: trained exclusively on clinical populations from their respective care settings; continuously refreshed as new patient data is documented; developed with clinical domain expertise; and validated against real-world outcomes monitored by internal clinical experts who review model outputs and feed findings back to data science teams for refinement.
That last element — clinical experts actively monitoring model outputs — is what separates a statistically derived risk score from a clinically trustworthy prediction. The model learns from what it gets right and wrong in the real world, not just from its original training data.
Download Mosai's clinical outcomes data and model architecture overview: Mosai | Success Story