Multi-Disease Detection of Diabetes and Breast Cancer Using The Doctor Framework

Machine Learning (ML) models using physiological signals and historical medical data to identify ailments are slowly gaining popularity. However, data about diseases becomes available incrementally over time and the behavior of diseases themselves could evolve over time. These factors require the ML models to be re-trained with the new data. The common approach of fine-tuning with new data can lead to catastrophic forgetting, where the model forgets what it learnt in prior iterations. Additionally, most disease detection models are trained for the detection of a single disease only. This increases the computational costs associated with ML model training and deployments, which would be critical when considering usage in cost-sensitive locales such as Africa. In this study, we tried to address these challenges by applying the DOCTOR framework, based on a Multi-head Deep Neural Network (DNN) and a Continual Learning, for the detection of diabetes and breast cancer. In a single-disease incremental scenario, the DOCTOR framework beat the baseline performance of the ideal joint scenario with average accuracy and f1 scores of 0.970 and 0.788 for diabetes, 0.875 and 0.805 for breast cancer. In the multi-disease setting, the DOCTOR framework showed the best performance for diabetes, even though it didn't better the ideal joint training scenario on breast cancer. Also, the results on the diabetes were even better than the single-disease scenario, which shows that such a multi-task learning scenario can help improve model generalization.