Artificial intelligence in orthopaedic recovery

Dealing with the timely challenge of helping medical teams increase their efficiency and identify crucial points in time where their intervention is needed, is the main objective of this research. Utilizing the data of orthopaedic patients, recovering from a knee/hip joint replacement surgery, we implement several Machine Learning (ML) models aimed at predicting patient pain levels and clinical interventions. We start with the construction of a flat dataset, holding patient daily reportings, consisting of over 400 different features which are collected from various sensors. In order to reduce noise and data dimensionality, we then apply different statistical techniques for both building new features from existing ones and selecting the most relevant features for each of our prediction tasks. Data cleansing and analysis methods, including feature distribution plots, Pearson correlation, data scaling and data imputation were used in order to create the narrowest, fullest and most appropriate dataset for each prediction task and ML model. Finally, we implement and train decision trees, Support Vector Machines (SVMs) and Recurrent Neural Networks (RNNs) for obtaining high performing classifiers for our objectives. The models built were carefully analyzed and tuned in order to extract maximum information about the most significant features and to be able to translate the results into recommendations for the clinical team. Based on separate blind test set evaluation, the different models (built for each research question) were able to predict the target with similar balanced accuracy (75% for manual intervention prediction and 85% for substantial pain prediction), however with a significantly better interpretability for the decision trees, making them a much more useful and insightful tool in this framework.