Abstract 18079: Forecasting Inpatient Cardiology Consultation and Transthoracic Echocardiogram Volumes

Introduction: The demand for inpatient cardiology consultation and transthoracic echocardiography (TTE) places a significant strain on hospital resources. There is limited understanding of the frequency and predictors of cardiology consultations and TTE in a real-world inpatient setting which hinders efforts to optimize staffing and workload expectations. Forecasting methods may shed light on these critical aspects of healthcare delivery.

Hypothesis: Forecasting methods may predict the daily volumes of inpatient cardiology consultations and TTEs in an academic medical center.

Methods: Our study examines the daily consultation and TTE ordering patterns using a dataset of over 26000 adult patients admitted to a teaching hospital in 2019 under “observation” or “inpatient” status. Using individual patient characteristics and clinical data, we sought to forecast the probability of a cardiology consult or TTE being ordered at the patient level and then aggregated these probabilities to predict daily volumes. The dataset was partitioned into a training cohort (80%) and a testing cohort (20%). Models tested included logistic regression, random forest, support vector machine, neural networks, gradient boosting machines, and hybrid models. Model performance was assessed using mean absolute error (MAE).

Results: In 2019, the average daily cardiology consult and TTE volumes were 2.8 (SD = 1.5) and 13 (SD = 4.5), respectively, in the hospital studied. No strong seasonal pattern was observed in the consultation and TTE orders. Utilizing individual-level forecasting followed by aggregation at a daily level resulted in reasonable daily estimates. Preliminary results indicate that the best-performing machine learning-based models for individual predictions yield MAE values ranging from 1.06 to 1.19 for consultations and from 3.18 to 3.62 for TTEs. These findings suggest that, on average, the predictions deviate by approximately 1 unit in absolute value for consultations and around 3 units in absolute value for TTEs from the true values.

Conclusion: Our findings suggest the feasibility of utilizing machine learning models for individual predictions to forecast and facilitate planning for hospital inpatient cardiology consultations and TTE volumes.