Background: Hospital discharge requires coordination among multiple disciplines and may feel chaotic as discharge approaches. The 48-Hour Discharge Prediction Tool (48DPT) is an AI-based system developed to predict clinical readiness for discharge 48 hours beforehand, with the aim of alerting the interdisciplinary team and prompting earlier completion of preparatory procedures. This study assessed 48DPT’s impact on workflow through observation of interdisciplinary rounds (IDR) and interviews with members of the care team. The Consolidated Framework for Implementation Research (CFIR) and Reach, Efficacy, Adoption, Implementation, and Maintenance (RE-AIM) frameworks were used to assess the impact of the tool (1,2) .

Methods: A research coordinator observed IDR on two medical units at a large urban academic hospital and assed 48DPT usage for patients who were “positive” (i.e., 48DPT indicated high likelihood of discharge). Observations noted whether the 48DPT was used by case management, and whether the tool and discharge planning were discussed during daily IDR. Semi-structured interviews with stakeholders – hospitalists, social workers, case managers, nurse managers, and unit medical directors – were conducted. Interview questions were developed based on RE-AIM and CFIR constructs applied to the 48DPT and discharge process. Direct content analysis assessed responses and mapped CFIR constructs onto RE-AIM domains.

Results: A total of 170 patent discussions for patients with positive 48DPT scores were observed over 30 dates. The 48DPT was routinely used (95.9%) by case managers and discharge was discussed for 97.3% of these patients. The 48DPT was rarely (9.3%) explicitly mentioned. When 48DPT was mentioned, clinicians agreed with the positive 48DPT score for 53% of patients. A total of 14 clinicians and members of the IDR team were interviewed. Results by RE-AIM Framework:• Reach: Varying use by social workers, case managers, and limited engagement by hospitalists highlighted potential barriers.• Efficacy: Confidence on 48DPT accuracy varied, with some stakeholders recognizing its potential to initiate discharge processes. Limited awareness among hospitalists hindered overall efficacy.• Adoption: Limited awareness among some stakeholders hindered full adoption, especially among less experienced clinicians.• Implementation: Lack of involvement in the tool’s development process hindered engagement. Support services positively impacted implementation by enhancing awareness and technical abilities.• Maintenance: The potential benefits outweighed the burden for most participants, primarily due to low burden. The primary barrier was limited feedback opportunities and inadequate resources.

Conclusions: Direct observation of IDR and qualitative interviews with stakeholders found that the 48DPT was widely utilized by case management and discharge discussions occurred for almost all patients with a positive 48DPT, though overall impact was small due to limited awareness and lack of involvement in the tool’s development and implementation, and limited feedback opportunities to the users. Integration of AI tools in clinical practice strategies should emphasize early engagement in the development and implementation processes, stakeholder education, and consistent clinician feedback.