YouDiagnose Predictive Model Pilot Study to Compare Expert Human vs. Machine Prediction Accuracy

YouDiagnose carried out a pilot study that seeks to compare the accuracy of expert human predictions with those made by a predictive model. Specifically, this study will analyse the prediction of diseases, cancer risk and the care need of patients being admitted or referred to the specialist service. The doctor was provided with a set of patient data, e.g., patient ID 9001 to 9050. The pilot consisted of 2 parts 1. Clinical prediction and 2. Model validation. In the first part, the doctor had to read the information in the clinical vignette and had to select one of the predetermined choices for the best clinical prediction (DP-1, DP-2, DP-3). Afterwards, he/she was asked to validate the models predictions (MP-1, MP-2, MP-3) and recommendations (R-1, R-2, R-3) for the same case. The study shows that MP has higher total accuracy (82.8%) compared to DP (50.6%). In predicting cancer, the MP method has higher sensitivity (100.0%) and positive predictive value (38.5%) compared to the DP method (90.0% sensitivity and 30.0% positive predictive value). The MP method also has higher specificity (78.9%) compared to the DP method (72.7%). Both methods have high negative predictive values (98.2% for DP and 100.0% for MP) with a p-value of 0.3705. The results of the current pilot study demonstrate the models potential, while also highlighting areas where further testing is needed in order to increase user confidence and improve the accuracy of diagnosis. Such testing could provide invaluable insights into how to maximize the value of the system in offering better frontline screening solutions e.g., triaging, clinical decision support, risk-based clinic booking system etc.