Responsible AI in Action: Planning through Implementation of a Mortality Model for Palliative Care

Importance

Interest in the use of prediction models to support referrals to palliative care is surging. Few high-performing models have been developed, implemented, and pilot-tested following responsible artificial intelligence (AI) principles and transparent reporting guidelines.

Objective

Collaborate with physicians and clinical informaticists to plan, develop, validate, implement, and pilot-test a high-performing, statistically fair prediction model—the Serious Illness Clinical Indicator (SICLI)—to support palliative care referrals following responsible AI principles.

Design

Cohort study of hospitalizations occurring from January 1st, 2022, to December 31st, 2022.

Setting

Kaiser Permanente Southern California

Participants

Patients 18+ years, not admitted to maternity or psychiatry departments and hospitalized for at least 36 hours.

Exposures

We augmented a commercial prediction model, the Epic End-of-Life-Care Index (EOLCI), by adding locally available predictors, including co-morbidity and laboratory acuity risk scores, nursing flow sheet and health care utilization data.

Main Outcomes and Measures

We predicted 12-month mortality risk using a split-sample design, least absolute shrinkage, and selection operator (LASSO) using logistic regression, and five-fold cross-validation. Performance was assessed via the Area Under the Receiver Operating Curve (AUC) and calibration plots. The final model was chosen by clinicians and informaticists for its balance of performance and parsimony. Responsible AI principles guided each development step.

Results

Twelve-month mortality rate among 133,043 hospitalizations of at least 36 hours was 23%. Patients who died were older (76±13 vs. 64±17 years) and had higher co-morbidity burden (Charlson, 4.49±3.14 vs. 2.06±2.44). SICLI was very well calibrated and outperformed the EOLCI (AUC of 0.87 (95% CI:0.86, 0.87) vs. 0.81, (95% CI: (0.80, 0.81)). SICLI’s high and very-high risk group (probability cut-off 0.6-<95 and >=0.95), produced positive predictive values (PPVs) of 72.0% and 94.0%, respectively. SICLI was implemented in KPSC’s Electronic Health Record. Post-implementation validation by a palliativist against the 12-month surprise question via review of 25 charts yielded a PPV of 100% and 80% for the very-high and high-risk groups. We report on model documentation, socialization, and governance.

Conclusion and Relevance

SICLI is an equitable, high-performing mortality prediction model that builds on and outperforms the EOLCI. It was implemented and pilot-tested in a large health care system.

Key points