Artificial Intelligence as a Screening Method for Consultation-Liaison Psychiatry in Patients Receiving Solid Organ Transplant

Abstract: Background/Significance: Proactive psychiatric consultation has been shown to reduce patient length of hospital stay compared to traditional consultation processes (Desan 2011). However, the screening processes have historically been questionnaires, chart review, or EMR alerts which can be time intensive and/or rudimentary. Via rapid analysis of large datasets, Artificial Intelligence (A.I.) may be able to provide an efficient and comprehensive screening method. Patients receiving solid organ transplants (SOT) are at increased risk for developing psychiatric disease (Dew, 2015) and will likely benefit from earlier identification of psychiatric symptoms. We aim to identify predictors of psychiatric consults and create an A.I. model of proactive consultation to identify at-risk hospitalized SOT patients.

Methods: A series of patient characteristics, including demographics, psychiatric diagnoses, psychotropic medications, and medical comorbidities were selected as input variables through literature review (Oldham, 2024) and an expert focus group of consultation-liaison psychiatrists. Following IRB approval, we retrospectively assessed for these predictors in all adult SOT (heart, lung, kidney, and liver) patients admitted between 2010-2023 to the University of Michigan University Hospital, a tertiary care transplant center. A multilayer perceptron neural network was implemented via TensorFlow API in R (ver. 4.3.3). Data imputation was utilized for continuous variables by supplying median values and model optimization was achieved using a weighted binary crossentropy function. 80% of the cohort was randomly sampled and utilized for model training while the remaining 20% was reserved for validation. The model was configured to output consult likelihood given input parameters of a specific encounter.

Results: A total of 88,436 inpatient encounters met inclusion criteria. Model accuracy was determined to be 71.6% after 100 epochs of training with a loss of 0.73. The model identified a mean probability of psychiatric consultation of 61.7% with a standard deviation of 2.5%. Overall model sensitivity was 89.0% with consult likelihood threshold at 0.5, 71.9% with a threshold of 0.75, and 97.5% with a threshold of 0.25.

Conclusion: With a sensitivity ranging between 71.9% and 97.5% based on parameter choice, the A.I. algorithm was successful at computationally qualifying the contribution of each predictor to the likelihood of requiring a psychiatry consultation during hospital admission. These data were consistent with our pre-study goal of a sensitivity of 80%, suggesting we successfully created an A.I. screening process for a transplant population. Future prospective evaluations of algorithm predictivity will provide insight into the integration of A.I. into transplant psychiatry practice.