OBJECTIVES: Late identification of clinical deterioration is a significant source of morbidity and mortality among hospitalized patients. Very few AI systems targeting early signatures of illness to prevent clinical deterioration have been implemented in practice and the economic impacts of these systems are not well understood. We assessed the hospital charges and costs of patients who were randomized to receive either a continuous visual risk score displayed versus standard of care in the acute care hospital setting.

METHODS: 10,422 patients were included in the pragmatic cluster-randomized controlled trial and both hospital charges and costs were obtained for each hospital admission. Generalized linear models were constructed to assess the relationship between the trial arm (AI-risk display versus standard of care) on cost outcomes in both the full cohort and in those at-risk of clinical deterioration. The randomization arms were analyzed as intention-to-treat.

RESULTS: There was evidence of differences in cost outcomes for the entire admission ranging from 10-22% differences by study arm favoring standard of care in both cohorts. In post hoc analysis among patients who had bed changes, we noted that there was a higher patient acuity among those transferred to an AI-risk display bed, thereby undermining the random nature of assignment in this real-world pragmatic design. This movement of sicker patients to intervention beds likely may have contributed to the findings.

CONCLUSIONS: Studying the economic consequences of AI-based risk scores and early warning systems remains challenging. There is a real possibility that early warning drives proactive action that results in earlier and longer interventions with improved patient and clinician outcomes. Further, costs and charges linked to the entire hospital stay might not be the most important economic marker of effectiveness. Developing nuanced simulation models can help health systems determine the cost impacts prior to implementation.