An Examination of Clinical Decision Support for Discharge Planning: Systematic Review, Simulation, and Natural Language Processing to Elucidate Referral Decision Making

Statement of the Problem: As healthcare data becomes increasingly prolific and older adult patient needs become more complex, there is opportunity for evidence-based technology such as clinical decision support systems (CDSS) to improve decision making at the point of care. Although CDSS for discharge planning is available, few published tools have been translated to new settings. Existing studies have not explored discordance between recommended and actual discharge disposition. Understanding the reasons why patients do not receive optimal post-acute care referrals is critical to improving the discharge planning process for older adults and their families. Methods: Three-paper dissertation examining CDSS. Paper 1 is a systematic review of studies with prediction models for post-acute care (PAC) destination. Paper 2 is a retrospective simulation of a discharge planning CDSS on electronic health record (EHR) data from two hospitals to examine differences in patient characteristics and 30-day readmission rates based on a CDSS recommendation among patients discharged home to self-care. Paper 3 is a natural language processing (NLP) study including retrospective analysis of narrative clinical notes to identify barriers to PAC among hospitalized older adults and create an NLP system to identify sentences containing negative patient preferences. Results: Most prediction models in the literature were developed for specific surgical populations using retrospective structured EHR data. Most models demonstrated high risk of bias and few published follow-up studies. In the simulation study, surgical patients identified by the CDSS as needing PAC but discharged home to self-care experienced adjusted 51.8% higher odds of 30-day readmission compared to those not identified. In the NLP study, the top three barriers were patient has a caregiver, negative preferences, and case management clinical reasoning. Most patients experienced multiple barriers. The negative preferences NLP system achieved an F1-Score of 0.916 using a deep learning model after internal validation. Conclusions: Future prediction modeling studies should follow TRIPOD guidelines to ensure rigorous reporting. Findings from the simulation and NLP studies suggest transportability of the CDSS to large urban academic health systems, especially among surgical patients. Incorporating natural language processing variables into CDSS tools may aid the identification of barriers to PAC