Artificial neural networks (ANN) have not been used in chronic subdural hematoma (CSDH) outcome prediction following surgery. We used two methods, namely logistic regression and ANN, to predict using eight variables CSDH outcome as assessed by the Glasgow outcome score (GOS) at discharge. We had 300 patients (213 men and 87 women) and potential predictors were age, sex, midline shift, intracranial air, hematoma density, hematoma thickness, brain atrophy, and Glasgow coma score (GCS). The dataset was randomly divided to three subsets: (1) training set (150 cases), (2) validation set (75 cases), and (3) test set (75 cases). The training and validation sets were combined for regression analysis. Patients aged 56.5 +/- 18.1 years and 228 (76.0%) of them had a favorable outcome. The prevalence of brain atrophy, intracranial air, midline shift, low GCS, thick hematoma, and hyperdense hematoma was 142 (47.3%), 156 (52.0%), 177 (59.0%), 82 (27.3%), 135 (45.0%), and 52 (17.3%), respectively. The regression model did not show an acceptable performance on the test set (area under the curve (AUC) = 0.594; 95% CI, 0.435-0.754; p = 0.250). It had a sensitivity of 69% and a specificity of 46%, and correctly classified 50.7% of cases. A four-layer 8-3-4-1 feedforward backpropagation ANN was then developed and trained. The ANN showed a remarkably superior performance compared to the regression model (AUC = 0.767; 95% CI, 0.652-0.882; p = 0.001). It had a sensitivity of 88% and a specificity of 68%, and correctly classified 218 (72.7%) cases. Considering that GOS strongly correlates with the risk of recurrence, the ANN model can also be used to predict the recurrence of CSDH.

Background This study develops machine learning (ML) algorithms that use preoperative‐only features to predict discharge‐to‐nonhome‐facility (DNHF) and length‐of‐stay (LOS) following complex head and neck surgeries. Methods Patients undergoing laryngectomy or composite tissue excision followed by free tissue transfer were extracted from the 2005 to 2017 NSQIP database. Results Among the 2786 included patients, DNHF and mean LOS were 421 (15.1%) and 11.7 ± 8.8 days. Four classification models for predicting DNHF with high specificities (range, 0.80‐0.84) were developed. The generalized linear and gradient boosting machine models performed best with receiver operating characteristic (ROC), accuracy, and negative predictive value (NPV) of 0.72‐0.73, 0.75‐0.76, and 0.88‐0.89. Four regression models for predicting LOS in days were developed, where all performed similarly with mean absolute error and root mean‐squared errors of 3.95‐3.98 and 5.14‐5.16. Both models were developed into an encrypted web‐based interface: https://uci-ent.shinyapps.io/head-neck/. Conclusion Novel and proof‐of‐concept ML models to predict DNHF and LOS were developed and published as web‐based interfaces.