BACKGROUND:Deep neural networks yield high predictive performance, yet obscure interpretability limits clinical applicability. We aimed to build an explainable deep neural network that elucidates factors associated with readmissions after rib fractures among nonelderly adults, termed DeepBackRib. We hypothesized that DeepBackRib could accurately predict readmissions and a game theoretic approach to elucidate how predictions are made would facilitate model explainability.
METHODS:We queried the 2017 National Readmissions Database for index hospitalization encounters of adults aged 18 to 64 years hospitalized with multiple rib fractures. The primary outcome was 3-month readmission(s). Study cohort was split 60-20-20 into trainingvalidation-test sets. Model input features included demographic/injury/index hospitalization characteristics and index hospitalization International Classification of Diseases, Tenth Revision, diagnosis codes. The seven-layer DeepBackRib comprised multipronged strategies to mitigate overfitting and was trained to optimize recall. Shapley additive explanation analysis identified the marginal contribution of each input feature for predicting readmissions.
RESULTS:A total of 20,260 patients met the inclusion criteria, among whom 11% (n = 2,185) experienced 3-month readmissions. Feature selection narrowed 3,164 candidate input features to 61, and DeepBackRib yielded 91%, 85%, and 82% recall on the training, validation, and test sets, respectively. Shapley additive explanation analysis quantified the marginal contribution of each input feature in determining DeepBackRib's predictions: underlying chronic obstructive pulmonary disease and long index hospitalization length of stay had positive associations with 3-month readmissions, while private primary payer and diagnosis of pneumothorax during index admission had negative associations.
CONCLUSION:We developed and internally validated a high-performing deep learning algorithm that elucidates factors associated with readmissions after rib fractures. Despite promising predictive performance, standalone deep learning algorithms are insufficient for clinical prediction tasks: a concerted effort is needed to ensure that clinical prediction algorithms remain explainable.