Physician Friendly Machine Learning
MetadataShow full item record
Artificial Intelligence (AI) and Machine Learning (ML) today has infiltrated almost all fields, helping catch patterns and make interesting conclusions from data. The surge in AI over the years can be attributed to two main facts: Increase in computation power of newer systems and the availability of data, both of which serve as seeds for building good prediction models. Medicine has slowly but steadily adopted AI over the years. Yet, traditional heuristic approaches and experience of physicians and doctors is heavily relied upon to this date. This thesis proposes two machine learning tools that can help doctors, physicians and medical researchers with their diagnosis and treatment procedures. Proposal 1 discusses Automatic Machine Learning (AutoML), which is a tool that helps automate the process of ML model building and fine-tuning, taking away the onus of fine tuning model parameters from the programmer. The resistance towards adoption of ML in the medical community stems from the idea that the tools and knowledge are only accessible to highly trained ML experts. This proposal is an attempt at breaking this age-old perception by proposing Auto-ML as a tool to build good ML models. The experiment done to substantiate this claim is to have a graduate student with sufficient experience in ML, manually build and fine-tune ML models on two publicly available cardiovascular disease prediction data-sets over a month and compare the performance with that of Auto-ML. The results prove that Auto-ML is capable of building models of similar accuracies in a time span of 30 minutes per data-set, with just a few lines of code. This should provide enough empirical evidence and encourage doctors to adopt ML as part of their research. Proposal 2 discusses the power of visualization of Convolutional Neural Networks (CNN) in performing classification tasks and how they help develop trust in doctors and medical researchers about model predictions. Gradient-weighted Class Activation Mapping (Grad-CAM) is used as a tool to generate localization maps indicating regions in the image that contributed to a certain prediction from the CNN, thereby instilling trust in medical professionals.