Predicting in-hospital mortality for stroke patients: Results differ across severity-measurement methods

Objective: To see whether severity-adjusted predictions of likelihoods of in-hospital death for stroke patients differed among severity measures. Methods: The study sample was 9,407 stroke patients from 94 hospitals, with 916 (9.7%) in-hospital deaths. Probability of death was calculated for each patient using logistic regression with age-sex and each of five severity measures as the independent variables: admission MedisGroups probability-of-death scores; scores based on 17 physiologic variables on admission; Disease Staging's probability-of-mortality model; the Severity Score of Patient Management Categories (PMCs); and the All Patient-Refined Diagnosis Groups (APR-DRGs). For each patient, the odds of death predicted by the severity measures were compared. The frequencies of seven clinical indicators of poor prognosis in stroke were examined for patients with very different odds of death predicted by different severity measures. Odds ratios were considered very different when the odds of death predicted by one severity measure was less than 0.5 or greater than 2.0 of that predicted by a second measure. Results: MedisGroups and the physiology scores predicted similar odds of death for 82.2% of the patients. MedisGroups and PMCs disagreed the most, with very different odds predicted for 61.6% of patients. Patients viewed as more severely ill by MedisGroups and the physiology score were more likely to have the clinical stroke findings than were patients seen as sicker by the other severity measures. This suggests that MedisGroups and the physiology score are more clinically credible. Conclusions: Some pairs of severity measures ranked over 60% of patients very differently by predicted probability of death. Studies of severity-adjusted stroke outcomes may produce different results depending on which severity measure is used for risk adjustment.