Clinical Decision Support: Interpretability and Applications in Patient Monitoring

Current clinical practice heavily relies on technology to support decisions. In particular, machine learning is ever more employed in decision support systems. This can be attributed to information overload, the fact that it becomes impossible for a clinician to take into account all available information. The drawback of this approach is that such decision support systems are often black boxes, yielding no insight into the reason of their decisions. In medical settings however, where trust and accountability are important issues, such systems should preferably be interpretable. In contrast, other domains almost fully rely on observation or subjective patient-reported questionnaires to quantify a medical situation. This is the case with assessment of physical capacity in patients suffering from chronic musculoskeletal conditions. With the advent of wearable technology this quantification can also be performed in an objective way to yield complementary information. This requires automatic activity recognition and assessment of the more challenging, but informative transitory activities, preferably in the home environment.

With that situation in mind, the research in this PhD focuses on two themes: interpretable decision support and patient monitoring using wearables. It also connects them by developing interpretable models for activity assessment. A first set of objectives is connected to interpretable classification systems. A second set is linked to the development of activity recognition and assessment algorithms within the scope of the SPARKLE project.

The first part of this work develops and improves two algorithms to extract interpretable medical scoring systems for binary classification from data. Jointly, they are called Interval Coded Scoring (ICS). The resulting models are piecewise constant over the intervals of selected variables. User interaction is possible at several points to steer the training process with expert knowledge and make a desired trade-off between performance and model simplicity. A first existing approach, lpICS, is extended to support interactions. It consists of a total variation regularized classification problem cast in the framework of Support Vector Machines, solved via Linear Programming. However, its time complexity and its uniqueness properties can be improved. Therefore, preselection is introduced as an interval-informed way to decrease the data dimensionality. Also, a second approach, enICS, is developed. It is based on elastic net cast as a dual Support Vector Machine and can be solved independently of the data dimensionality. However, it does not support variable interactions. Both algorithms are validated on several publicly available biomedical datasets, proving a performance that is similar to several standard machine learning algorithms. ICS is implemented in a Matlab toolbox offering an interactive interface for model setup, training, risk estimation and visualization.

The second part designs and implements two pattern-based approaches for activity recognition using data from a single accelerometer mounted on the upper arm. They are both based on the assumption that pattern matching is more suitable for recognition of transitory activities, whereas sliding windows with statistical features are more prevalent in literature for repetitive activities of longer duration. One approach improves pattern matching recognition via the combination of dynamic time warping and common statistical features. It consists of an automatic segmentation approach followed by a multilayer random forest classification framework to reject false positives belonging to the rejection class. The other approach exploits the multilinear structure of the data. Higher Order Discriminant Analysis is applied on pattern-matched data to extract discriminatory features from the automatically segmented data. Both methods have been validated on acquired patient datasets in a protocol setting suitable for a home environment. They are proven to significantly outperform approaches based on either simple pattern matching or statistical features alone.
Moreover, the second part also studies assessment of activity recognition. Two informative parameter extraction methods are developed. Furthermore, three different ways to implement an objective complement to the BASFI questionnaire, via classification with ICS or regression, are presented. To end, ICS is applied to create a scoring system for comparison with the current standard in evaluation of disease activity.