Infrastructure for Mortality Risk Prediction

Mortality risk prediction requires explicit, findable documentation of trigger thresholds for palliative care referrals, goals-of-care conversation protocols, and code status documentation requirements. TJC and CMS conditions of participation formalize advance directive policies. The ML model needs clearly documented criteria for when a mortality risk score triggers a palliative care consult vs. a physician notification vs. a care plan update—these clinical decision rules must be findable and current, not residing in individual physicians' judgment.

Mortality risk prediction requires systematic capture of frailty assessments, functional status scores, advance directive status, code status, prior utilization patterns, and vital signs across encounters via structured EHR templates. Regulatory requirements for advance directive documentation and structured discharge summaries enforce baseline systematic capture. The model needs longitudinal patient data captured consistently across inpatient, outpatient, and ED encounters to compute 30/90-day post-discharge mortality risk.

Mortality prediction requires consistent schema across all clinical records: ICD-10 diagnosis codes, CCI comorbidity index components, functional status assessments (Katz, Barthel), vital sign trends, and lab values must share uniform field definitions. The model computes survival probability from structured feature vectors assembled from these consistently defined data elements. Without consistent schema, comorbidity burden calculations produce incomparable risk scores across patient populations.

Mortality risk prediction requires API access to EHR clinical data, prior utilization records, advance directive repositories, and palliative care consultation workflows. The model must surface risk scores in the attending physician's EHR workflow and trigger palliative care team worklists—not in a standalone analytics portal. API access to EHR and care management systems enables the model's output to reach clinicians at point of care during rounds.

Mortality risk model maintenance follows a periodic review cadence aligned with annual CMS reporting requirements and major clinical guideline updates. The model's risk factors—diagnosis weights, frailty indicators, utilization patterns—evolve slowly enough that scheduled annual or semi-annual recalibration is operationally acceptable for this capability. Unlike sepsis or ADE detection, mortality prediction supports longer-horizon care planning rather than immediate clinical response, reducing the urgency of real-time model currency.

Mortality risk prediction is primarily bounded to EHR-resident clinical data: diagnoses, labs, vitals, functional status, and advance directives. Point-to-point integrations between the EHR and palliative care consultation system, plus a link to the advance directive registry, are sufficient for the core workflow. Population-level mortality trending requires only EHR data extraction. Full API-based multi-system integration isn't required because the capability doesn't depend on real-time cross-system context assembly.