Prospective Study of a Free-text Diagnosis Prediction Algorithm for Appendicitis in the Emergency Department

Brief Summary

Detailed Description

Developing machine learning models that have a strong prediction power for diagnosis of appendicitis from physician entered free text input can improve diagnostic accuracy of doctors. It also offers the possibility of using prediction algorithms to improve routine clinical care. In the future, multiple machine learning models can be combined to increase prediction accuracy and prediction algorithms can be extended to other diagnoses. 18,000 cases of emergency department presentations over 10 years were used as a training and validation dataset. To develop the appendicitis prediction model, deep learning neural networks with a customized medical ontology were used. The diagnostic accuracy of the model is expressed as sensitivity (recall), specificity and F1 score (harmonic mean). The developed diagnosis predictive model shows high sensitivity (86.3%), specificity (91.9%) and F1 score (88.8) in diagnosing appendicitis from patients presenting with abdominal pain. The predictive model algorithm will also highlight words in the free text (entered by the attending physician) that it assigns higher probability for predicting an outcome. The doctors will be instructed to provide a percentage likelihood of appendicitis based on the clinical presentation and any available laboratory investigations. The doctor is then shown the prediction of the algorithm as well as the highlighted words for the patient entered. He/she must then provide another prediction of the likelihood of appendicitis after seeing the algorithm generated prediction. The aim is to evaluate the performance of the algorithm and to assess if usage of the algorithm is able to help emergency doctors improve their diagnosis of appendicitis. The prediction results will be tabulated to assess accuracy of the algorithm, doctors before algorithm input and doctors after receiving algorithm input. The accuracy will be expressed as sensitivity, specificity, accuracy, positive prediction value, F1 score and F0.5 score. Approximately 100 emergency doctors will be recruited over the course of 1 year as participants in the study. The doctors will be split randomly assigned to two groups - the algorithm arm and the no algorithm arm. The randomization will be by time (weekly) using variable block randomization of 4 and 6. The patients will be followed up for the final discharge diagnoses.

Primary Outcomes