Robust Smartphone Placement Classification To Enhance Digital Biomarker Reliability

Aim Smartphone-based digital movement biomarkers are gaining popularity in neurodegenerative diseases. In this context, phone placement critically affects the accuracy of the algorithms used to derive gait metric from sensor data under unsupervised conditions. This thesis addresses this challenge by developing a robust classification pipeline to automatically detect phone location. The ultimate objective is to enhance the reliability of smartphone-derived features, particularly in real-world conditions where phone placement is uncontrolled. Methods The classifier was developed and tested using three datasets. The lab_testing dataset (n=83 healthy adults) was used for training and validation. Participants performed a 30-second walk test wearing smartphones in three predefined positions: front pockets (FP), back pockets (BP), and a waist pouch positioned at the belly (BB). Inertial sensor data (accelerometer and gyroscope) were preprocessed and segmented into 2.5s overlapping windows. A total of 465 features per window were extracted from time, frequency, and energy domains. A multiclass gradient-boosted random forest (RXGBoost) was trained using 100 random hyperparameter configurations, and each configuration was evaluated with a 10-group-user-out cross validation. Two external datasets were used for testing the model’s robustness. The baseline dataset (n=61 people with Multiple Sclerosis, EDSS 0–3) shared the same protocol as the training data and was used to evaluate population-wise generalization. The Torino dataset (n=9 healthy adults) included 6-minute walk tests and a different phone placement scheme, with one phone in the FP and another one in a belt at the lower back (considered BB). This dataset tested robustness to changes in the experimental protocol. To handle uncertainty and avoid unreliable predictions, a margin-based threshold flags ambiguous case as 'Unknown’. Results On the lab_testing dataset, the classifier achieved optimal performance (F1-score = 0.996 ± 0.007, Matthews Correlation Coefficient, MCC = 0.994 ± 0.010, Accuracy = 0.997 ± 0.004, ROC-AUC = 0.9999 ± 0.0002). On the baseline dataset, the model maintained excellent results: F1-score = 0.941, MCC = 0.837, Accuracy = 0.968, ROC-AUC = 0.997. On the Torino dataset, despite the difference in protocol and task duration, the model still performed robustly (F1-score = 0.923, MCC = 0.892, Accuracy = 0.944, ROC-AUC = 1). The use of uncertainty management did not lead to an improvement in classification performance: results on the baseline dataset remained stable, while a slight decrease was observed on the Torino dataset (F1-score = 0.875, MCC = 0.842, Accuracy = 0.917, ROC-AUC = 1). However, this approach allows for the exclusion of predictions that may be correct but for which the model is uncertain, ensuring that retained predictions are associated with high confidence. Conclusion This work demonstrated that a machine learning classifier trained on healthy individuals can generalize well to patients with MS and across heterogeneous data collection protocols. The use of a classification pipeline combined with uncertainty management strategies improves the interpretability and reliability of sensor-derived features in real-world unsupervised settings. This methodological advancement supports the development of robust digital biomarkers and strengthens the potential of smartphone-based remote monitoring in clinical research and personalized disease management.