O11.3. DEEP LEARNING-BASED HUMAN ACTIVITY RECOGNITION FOR CONTINUOUS ACTIVITY AND GESTURE MONITORING FOR SCHIZOPHRENIA PATIENTS WITH NEGATIVE SYMPTOMS

• Published in: Schizophrenia bulletin Vol. 45; no. Supplement_2; pp. S194 - S195
• Main Authors: Umbricht, Daniel, Cheng, Wei-Yi, Lipsmeier, Florian, Bamdadian, Atieh, Tamburri, Paul, Lindenmann, Michael
• Format: Journal Article
• Language: English
• Published: US Oxford University Press 09.04.2019
• Subjects: Clinical trials; Deep learning; Oral Abstracts; Patients; Schizophrenia; Sensors
• ISSN: 0586-7614; 1745-1701
• DOI: 10.1093/schbul/sbz021.259

Abstract Background Quantified evaluation and continuous monitoring of behaviour and symptoms associated with negative symptoms of schizophrenia has been a challenge in clinical trials. Wearable sensor technology provides new opportunities to tackle these problems. In addition, the recent breakthrough in Deep Learning offers a novel approach for accurately inferring subjects’ behaviour through wearable sensor signals. The combination of the two could make continuous remote monitoring of schizophrenia patients in a clinical trial setting a reality. However, to connect the sensor data with clinical observations and measures of motivated behaviour still requires development of advanced analytics and real-world validations. Methods In a 3-way cross-over proof of mechanism (POM) study of a compound targeting negative symptoms of schizophrenia (Roche study BP29904) 33 patients with moderate negative symptoms were recruited (30 males; Mean age 36.6±7 y; BNSS total score =36.0 ±3.6; PANSS NSFS = 22.8 ±1.4; PANSS PSFS = 19.4 ±1.8). In 31 patients actigraphy data were recorded for 15 weeks. Motivated behaviour was assessed with an effort choice task. We trained a 9-layer convolutional recurrent neural network using two public data sets containing wrist-worn acceleration data to infer the subjects’ activities. The trained human activity recognition (HAR) achieved more than 95% of accuracy to separate the ambulatory (walking, stairs, cycling, jogging) activities from stationary activities (sitting, standing, lying down, hand work). We applied the HAR model to the actigraphy data of the POM study and calculated active time ratios for each patient. To infer gesture events, we extracted the time spans where the patients were stationary, while the standard deviation of the accelerometer signals exceeded an empirically-defined threshold of 0.1 m/s2. From derived gesture events we generated number of gesture events and the gesture power. For correlational analyses, actigraphy data obtained over the last 5 days of the placebo period and assessments at the end of the placebo period were used. Correlations were calculated using Spearman correlation and P values were corrected using Banjamini-Hochberg procedure. Results The patient adherence rate was high: average monitoring time per patient was 1,859 hours. The active time ratio was associated with the % high effort choice (r = 0.58, P = 0.002). The median daily gesture counts were negatively correlated with BNSS total score (r = -0.44, P = 0.03). Specifically, diminished expression sub-scores correlated negatively with both median daily gesture counts and median gesture powers (r = -0.42, P = 0.03 in both cases). We observed negative correlation between the changes in the diminished expression sub-score and the changes in median gesture counts between placebo period versus the dosage period (r = -0.47, P = 0.07 at low dose vs PBO, r = -0.34, P = 0.21 at high dose vs PBO), as well as changes in median gesture power (r = -0.39, P = 0.12 at low dose vs PBO, r = -0.7, P = 0.01 at high dose vs PBO). Discussion These analyses demonstrate the feasibility to the use of wrist-worn actigraphy for continuously monitoring clinically-relevant behaviour in a clinical trial setting. Associations with key dimensions of negative symptoms support their validity. The activity and gesture features derived from human activity recognition model shows promise for future clinical practice and drug development process. While we have seen encouraging correlations, a further validation study with more patients is still required to establish conclusive statistical evidence.