Abstract-This paper presents a technique for automatically

Radar systems can be used to perform human activity recognition in a privacy preserving manner. Deep Neural Networks are able to effectively process the complex radar data and make predictions. Often these networks are large and do not scale well when processing a large amount of radar streams at once, for example when monitoring multiple rooms in a hospital. This work proposes Bayesian Split Bidirectional Recurrent Neural Network for Human Activity Recognition. Using this technique the processing of data is split in two parts, one part on-premise (low-power, low-cost device), and the other off-premise (high power device). The proposed approach leverages the power of the off-premise device to quantify its uncertainty, and to gain more information on its epistemic and its aleatoric parts. Results indicate the proposed approach is able to correctly identify parts of a prediction that either need more training data for better predictions (epistemic uncertainty), or are inherently hard to classify by the model (aleatoric uncertainty).

Background: Disability progression is a key milestone in the disease evolution of people with multiple sclerosis (PwMS). Prediction models of disability progression have not yet reached the level of trust needed to be adopted in the clinic. A common benchmark to assess model development in multiple sclerosis is also currently lacking. Methods: Data of adult PwMS with a follow-up of at least three years from 146 MS centers, spread over 40 countries and collected by the MSBase consortium was used. With basic inclusion criteria for quality requirements, it represents a total of 15,240 PwMS. External validation was performed and repeated five times to assess the significance of the results. TRIPOD guidelines were followed. Confirmed disability progression after two years was predicted, with a confirmation window of six months. Only routinely collected variables were used such as the expended disability status scale, treatment, relapse information, and MS course. To learn the probability of disability progression, state-of-the-art machine learning models were investigated. The discrimination performance of the models is evaluated on their area under the receiver operator curve (ROC-AUC) and under the precision recall curve (AUC-PR), and their calibration via the Brier score and the expected calibration error. Findings: A temporal attention model was the best model. It achieved a ROC-AUC of 0.71 ± 0.01, an AUC-PR of $0.26 ± 0.02$, a Brier score of $0.1 ± 0.01$ and an expected calibration error of $0.07 ± 0.04$. The history of disability progression is more predictive for future disability progression than the treatment or relapses. Interpretation: Good discrimination and calibration performance on an external validation set is achieved, using only routinely collected variables. This makes these models ready for a clinical impact study. % in MS centers participating in MSBase. All our preprocessing and model code is available at https://gitlab.com/edebrouwer/ms_benchmark, making this task an ideal benchmark for predicting disability progression in MS.

Monitoring patients in hospitals or care homes using radars is an interesting problem with life-saving applications. Today deep neural networks are employed for patient monitoring, but these do not provide uncertainties, nor do they consider the asymmetry in the real life cost of misclassifying different activities. In this work we use Bayesian Neural Networks that provide uncertainty on their predictions. We combine these models with a self-defined utility function to obtain tailored predictions that are more conservative for classes where misclassifications come at a higher risk or cost. We show that Bayesian neural networks are more robust, and generalize better on radar human-activity images than deterministic ones, and that they are able to reduce the cost of misclassifications in a realistic example setting by 37% compared to approaches from literature.