Background: The observed consultation length at specialty clinics, such as cardiology care, is represented by two underlying groups one with zero service time due to patient no-shows, and the other characterized by positive values with high variance. This inconstancy affects the scheduler's ability to accurately estimate consultation length, which, in turn, hinders effective utilization of the clinic's resources and timely access to care. The objectives of this study were to: (i) predict the consultation length by accounting for its semicontinuous nature (i.e., zero in case of no-shows and positive otherwise), using machine learning (ML) algorithms, (ii) identify important features for predicting no-shows and non-zero consultation length, and (iii) assess the impact of integrating the ML-based prediction with the appointment scheduling system. Methods: We used two-years of data extracted from the electronic medical records of a cardiology clinic. By leveraging 16 predictors pertaining to the patient, appointment, and doctor, a two-part ML-based approach was developed to handle the semicontinuous consultation length. Supervised classification models were employed to predict no-shows (i.e., categorize the consultation length as zero or positive), and regression algorithms were developed for estimating non-zero consultation lengths. Three algorithms, namely, random forests, stochastic gradient boosting, and deep neural networks, were individually employed for both no-show classification and positive consultation length prediction. Finally, the best performing classification and regression models were combined to establish the complete two-part model, and its prediction error on new data is benchmarked against the clinic's current performance. The evaluation metrics for classification models were area under the receiver operating characteristic curve (AUC-ROC) and area under the precision-recall curve (AUC-PR). The prediction performance of regression algorithms was evaluated by mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE). A simulation modeling approach was adopted to ascertain the effectiveness of using ML-based prediction for scheduling decisions as opposed to the clinic's current strategy. Results: Among the classification models tested, stochastic gradient boosted classification tree (SGBCT) demonstrated best performance (AUC-ROC = 0.85, AUC-PR = 0.64). For positive consultation length prediction, deep neural network regressor (DNNR) resulted in lowest prediction error (MAE = 8.55, RMSE = 6.88, MAPE = 12.24). The complete two-part model (SGBCT + DNNR) outperformed the clinic's approach to consultation length estimation by achieving 50 % and 52 % reduction in RMSE and MAE, respectively. Further adopting it for appointment scheduling could reduce the patient waiting time and doctor idle time by 56 % and 52 %, respectively. Besides, several clinical insights, along with critical features for no-show and consultation length prediction, were also identified from our proof-of-concept study. Conclusion: This study demonstrates that routine clinical tasks such as estimation of consultation length and noshows can be accurately predicted using ML algorithms, and subsequently integrated into the clinical scheduling system to improve resource utilization and reduce patient waiting time.
