Explainable machine learning for real-time deterioration alert prediction to guide pre-emptive treatment

Daily prediction of next-day disease severity in critically ill patients: A prospective cohort study

BACKGROUND

Predictive models for intensive care unit (ICU) patients mainly focus on mortality, but short-term disease severity is more relevant for day-to-day decision-making.

AIM

The aim of this study was to develop and validate a daily prediction model for next-day disease severity in ICU patients.

METHODS

Data from the Simple Intensive Care Studies-II prospective cohort study of acutely admitted critically ill adults, including data collected during the first 7 days of admission such as Sequential Organ Failure Assessment (SOFA) score-related measurements, were used to fit a mixed-effects logistic regression model for next-day deterioration. Deterioration was defined as a decline in the total (≥2) and organ-specific (≥1) SOFA scores. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis.

RESULTS

A total of 1009 patients were included. The final predictive model for overall next-day deterioration included six predictors (the total SOFA score on the current day, the minimum value of arterial pH, Glasgow Coma Scale score, mean arterial blood pressure, mechanical ventilation, and its effect differing between the first and subsequent ICU days). The model achieved an AUC of 0.74 (95% confidence interval: 0.70-0.78). In the decision curve analysis, within probability thresholds of 0.2-0.5, the model showed a higher net benefit than did strategies of treating all patients or treating no patients. Organ-specific prediction models for next-day deterioration in respiration, cardiovascular, and renal function showed slightly better performance than the overall model (AUCs: 0.79, 0.79, and 0.81, respectively).

CONCLUSIONS

Daily prediction models can predict next-day disease severity in overall, respiration, cardiovascular, and renal function amongst ICU patients. They offer clinical benefits within a range of probability thresholds and could support decision-making for ICU physicians.

Expanding Critical Care Delivery beyond the Intensive Care Unit: Determining the Design and Implementation Needs for a Tele-Critical Care Consultation Service

BACKGROUND

 Unplanned intensive care unit (ICU) admissions from medical/surgical floors and increased boarding times of ICU patients in the emergency department (ED) are common; approximately half of these are associated with adverse events. We explore the potential role of a tele-critical care consult service (TC3) in managing critically ill patients outside of the ICU and potentially preventing low-acuity unplanned admissions and also investigate its design and implementation needs.

METHODS

 We conducted a qualitative study involving general observations of the units, shadowing of clinicians during patient transfers, and interviews with clinicians from the ED, medical/surgical floor units and their ICU counterparts, tele-ICU, and the rapid response team at a large academic medical center in St. Louis, Missouri, United States. We used a hybrid thematic analysis approach supported by open and structured coding using the Consolidated Framework for Implementation Research (CFIR).

RESULTS

 Over 165 hours of observations/shadowing and 26 clinician interviews were conducted. Our findings suggest that a tele-critical care consult (TC3) service can prevent avoidable, lower acuity ICU admissions by offering a second set of eyes via remote monitoring and providing guidance to bedside and rapid response teams in the care delivery of these patients on the floor/ED. CFIR-informed enablers impacting the successful implementation of the TC3 service included the optional and on-demand features of the TC3 service, around-the-clock availability, and continuous access to trained critical care clinicians for avoidable lower acuity (ALA) patients outside of the ICU, familiarity with tele-ICU staff, and a willingness to try alternative patient risk mitigation strategies for ALA patients (suggested by TC3), before transferring all unplanned admissions to ICUs. Conversely, the CFIR-informed barriers to implementation included a desire to uphold physician autonomy by floor/ED clinicians, potential role conflicts with rapid response teams, additional workload for floor/ED nurses, concerns about obstructing unavoidable, higher acuity admissions, and discomfort with audio-visual tools. To amplify these potential enablers and mitigate potential barriers to TC3 implementation, informed by this study, we propose two key characteristics-essential for extending the delivery of critical care services beyond the ICU-underlying a telemedicine critical care consultation model including its virtual footprint and on-demand and optional service features.

CONCLUSION

 Tele-critical care represents an innovative strategy for delivering safe and high-quality critical care services to lower acuity borderline patients outside the ICU setting.