Erratum: Machine learning-based prediction of health outcomes in pediatric organ transplantation recipients (JAMIA Open (2021) 4:1 (ooab008) DOI: 10.1093/jamiaopen/ooab008)

This is a correction to: JAMIA Open, Volume 4, Issue 1, January 2021, ooab008, https://doi.org/10.1093/jamiaopen/ooab008 Due to an error in the implementation code of cross-validation part of the deep learning models, the performance of the models was not estimated correctly. This happened because a new model was not properly built inside the K-fold loop, causing the parameters learned in the previous fold to be carried over to the next fold. We apologize for this error and hereby make the following corrections to the paper. Details of the corrections: 1. Michael O. Killian and Zhe He are co-corresponding authors. 2. Page 1 (Results section of the abstract): Error: "DL models generally outperformed traditional ML models across organ types and prediction windows with area under the receiver operating characteristic curve values ranging from 0.750 to 0.851." Correction: "DL models did not outperform traditional ML models across organ types and prediction windows with area under the receiver operating characteristic curve values ranging from 0.50 to 0.593." Error: "Results demonstrate the utility of DL modeling for health outcome prediction with pediatric patients, and its use represents an important development in the prediction of post-Transplant outcomes in pediatric transplantation compared to prior research" Correction: "Results showed that deep learning models did not yield superior performance in comparison to models using traditional machine learning methods. However, the potential utility of deep learning modeling for health outcome prediction with pediatric patients in the presence of a large number of samples warrants further examination." 3. Page 5 (Results section of the main text): Error: "DL models generally outperformed all traditional ML models across organ types and prediction windows (Figure 2). The average performance of DL models across all three prediction windows for kidney organ type was 0.80560 6 0.028, for liver organ type was 0.78960 6 0.010, and for was 0.812606 0.054." Correction: "DL models did not outperform traditional ML models across organ types and prediction windows (Figure 2). The average AUROC of DL models across all three prediction windows for kidney transplantation was 0.5455 6 0.018, for liver transplantation was 0.5358 6 0.016, and for heart transplantation was 0.5528 6 0.047." Error: "The best performing DL model across all organ types and prediction windows was for 3-year hospitalization after heart transplantation with an AUROC of 0.851" Correction: "The best performing DL model across all organ types and prediction windows was for 3-year hospitalization after heart transplantation with an AUROC of 0.593." 4. Page 5 (Discussion section of the main text): Error: "Results from the current analyses demonstrate the promise of DL models in the prediction of adverse medical events during posttransplant care of pediatric liver, kidney, and heart transplant patients.". . . "Accuracy of DL models here outperformed a recent examination of pediatric liver transplantation using the Studies of Pediatric Liver Transplantation data and a RF decision tree approach to ML[15]. Importantly, DL models offer increased predictive utility through the examination and evaluation of complex relationships among variables as important pathways (e.g., neurons activation) for improving outcome prediction [42]." Correction: "Based on the current analyses, although deep learning did not outperform traditional methods, its potential predictive utility should not be underestimated. The datasets in this study were small with a mixed of numeric and categorical variables. In a follow-up study on whole UNOS data, we are reexamining the performance of deep learning models in comparison to traditional ML methods in the presence of larger number of samples". (Table presented).