Machine Learning Models to Predict Diagnosis and Surgical Outcomes in Otolaryngology
Background: Despite the rapid evolvement of machine learning (ML) applications within the medical literature, there exist a paucity of knowledge regarding diagnosis, decision-making, and predicting surgical outcomes in otolaryngology using ML models. This thesis aims to 1) Construct ML models that use pre-operative-only inputs to predict length of stay (LOS) and discharge to non-home facility (DNHF) following complex head and neck (HN) surgeries, 2) Utilize deep learning and a large number of input variables to predict short-term adverse events following vestibular schwannoma (VS) surgery, 3) Construct preliminary binary (normal vs. abnormal) and multiclass (normal vs. acute otitis media vs. otitis externa vs. chronic suppurative otitis media vs. cerumen) image classification models of otoscopic images, and 4) Publish the best-performing models as public web-based interfaces.
Methods: To develop novel ML models that predict various post-operative outcomes, the 2005-2017 National Surgical Quality Improvement Program database was utilized to extract subjects undergoing major HN surgery (N=2667) and VS surgery (N=1783). Datasets were randomly stratified into training and testing sets using an 80:20 ratio with k-fold cross-validation used for training. To develop the novel otoscopic image classification model utilizing Inception-Resnet-V2 networks, 400 publicly available and labeled otoscopic images were obtained.
Results: Four classification models for predicting DNHF were developed with high specificities (range 0.80-0.84), where the generalized linear model and gradient boosting machine models outperformed artificial neural network and random forest models 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, respectively. Four regression models were developed for predicting LOS in days, where all performed similarly with mean absolute error and root mean squared errors of 3.95-3.98 and 5.14-5.16, respectively. The DNHF and LOS models were developed into a web-based interface: https://uci-ent.shinyapps.io/head-neck/. Using pre-, peri-, and post-operative inputs, three deep-learning models to predict unplanned reoperation, surgical complications, and medical complications following VS surgery were developed with ROC of 0.74, 0.70, and 0.83, and accuracy of 0.91, 0.71, and 0.92, respectively. Lastly, a preliminary image classification model with ROC, accuracy, and NPV of 0.89, 0.77, and 0.74 for binary classification, and the capability to suggest the diagnosis among abnormal images, was developed: https://headneckml.com/tm/.
Conclusion: Novel ML models to predict DNHF or LOS following complex HN surgery, reoperation or complications following VS surgery, and abnormalities on otoscopic imaging were successfully developed. Publishing such models as interactive web-based interfaces will help advance this frontier by providing opportunities for examination/validation and practical benefit to clinicians.