Enhanced Prediction of Osteoporosis with Improved  Contextual Accuracy Utilizing Expanded  Classifications and Explainable AI

Many classification models shown in medical research achieve high statistical accuracy, yet are not examined for their contextual accuracy, lack of which can lead to misleading predictions that do not align with human reasoning or the science involved. This will cause distrust and alienate those unfamiliar with deep learning and training models, even if they understood the medical condition being predicted. To address this, explainable AI can evaluate a model’s contextual accuracy by assessing how reasonable its decision-making process appears to be. In this paper utilization of class augmentation as a method to improve contextual accuracy in classification models is explored. This forces the model to rely on finer details within the image, ultimately enhancing its reasoning ability. This is explored within the context of using AI models to diagnose patients with Osteoporosis using X-ray images instead of DXA. Osteoporosis is a silent, bone density disorder that needs to be diagnosed as early as possible to avoid many complications, since in serious cases, even a simple sneeze can fracture a bone. Understanding why a model would classify a patient as healthy or having osteoporosis would help physicians judge this decision better. Explainable AI is used to evaluate this decision making, providing insights into how well the model's decision making aligns with human logic. To validate the findings, two different x-ray datasets were used, where the images within each dataset appeared visually similar to one another within the dataset itself. Results show that utilizing class augmentation when training a model before later training it for a specific use case with a smaller number of classes can give better explainability to the resulting model. In a medical setting, models can be forced to learn better clinical markers during training. Additionally, this had consistently lowered the number of false positives.