# Anomaly detection to predict relapse risk in schizophrenia Philip Henson 1,2, Ryan D’Mello2, Aditya Vaidyam2, Matcheri Keshavan2 and John Torous 3 [[Anomaly detection to predict relapse risk in schizophrenia (1).pdf]] ## My Notes - Similar ideas to HOPE-s to use anomaly detection to predict relapses. - 15 Patients. 5 Relapses. 3 had false positives detected. 2 didn't have enough data to predict relapse - I am curious about what data of patient who did not relapse, how does these data look like? What value does it have? ## Methods - 17 patients with schizophrenia. Use Beiwe App. 3 months. - Collect GPS, Accelerometer, call, text logs. Screen on and off time. phone charging status. - We defined clinical relapse as either psychiatric hospitalization or an increase in the level of psychiatric care, such as increase in the frequency of clinic visits or referral to a partial or outpatient hospital program - These in-app surveys contained questions designed to measure anxiety, depression, sleep quality, psychosis, the warning symptoms scale, and whether or not the subjects were taking their medication. - With this wide range of mobility features, sociability features, and clinical outcomes being recorded up to 3 months on a daily basis for each patient, ==trends in patient behavior can be established by treating these features as a multivariate time series. Anomalous breaks from a patient’s usual trend in their self- reported outcomes, sociability, or mobility may be indicative of broader behavioral changes and could precede adverse events, such as relapse. == ## Results We investigate the rate at which significant anomalies occur across the 15 patients in the sample (Fig. 1). On average, there were 1.8 significant anomalies in mobility detected per month across the sample, 1.7 significant anomalies in sociability detected per month across the sample, and 1.4 significant anomalies in self- report of clinical outcomes detected per month across the sample. In all three cases, the variability in this rate varied greatly across subjects, with some subjects having over twice the rate of anomalies detected as some others. If anomaly detection is to be used to prompt interventions in the future, the frequency of anomalies detected should ideally match the frequency of actual relapses without having too many false positives. Given the rarity of relapse and the small sample size used in this pilot study, sensitivity and specificity cannot be accurately estimated from our data. - 15 patients tracked. 5 relapsed. - 3 had sufficient data to fest for presence of behavioral anomalies. - 2 subjects does not have enough data because uninstalled 3 weeks before readmission, 1 uninstalled one week before readmission. Of the three subjects who experienced relapse with sufficient data collected to perform anomaly detection analyses, ==the rate of anomalies detected in the 2 weeks prior to relapse was 71% higher than the rate of anomalies detected in dates further away from relapse (Table 1)==. Due to the relative rarity of relapse events in our sample, a 2-week window was selected instead of 1 week in order to allow for sufficient data in the calculation of anomaly rates. - This 3 subjects - 71% higher rate of anomalies, in dates further away from relapse (what does that mean?) Of the patients with relapse and hospitalization, the first patient had 48 days of follow-up, 37 of which were collected prior to hospitalization; the second patient had 53 days of follow-up, 48 of which were collected prior to hospitalization; and the third patient had 79 days of follow-up, 41 of which were collected prior to hospitalization. ==In this third patient, a week prior to relapse and hospitalization there were significant anomalies in mobility, sociability, and self-reported clinical outcomes == **- Which is the 3 patients with enough data to talk about.** Despite adhering to the study and having reliable smartphone data collection for the first month, no data were collected for 2 days prior to the significant behavioral anomalies, which could also be related to the patient’s regression. On the anomalous day, 9 days before hospitalization, ==the subject spent only 6 h and 51 min out of the 24-h period at home, far less time than usual.== Also, in contrast with most previous days during which the ==subject communicated with no one through their phone’s SMS system or calls, the subject made four phone calls with the same phone number on the anomalous day.== Additionally, on the anomalous day the ==patient reported trouble with moderately strong anxiety (2 on a 0–3 scale), moderately strong depression (2 on a 0–3 scale), moderately high on the Warning Symptoms Scale (2.25 on a 0–3 scale), extreme sleeping problems (3 on a 0–3 scale), and extreme levels of psychosis (3 on a 0–3 scale).== These scores represent an increase across the board in all categories when compared to self-report from a week before the anomaly. **It is the similar curiosities, about making sense of the data collected. e.g** _- How come patient spent less time at home? Any reasons?_ _- How come patient used the phone less? any other reasons?_ _- The patient called someone 4 times? OK, so what? Who did patient call?_ _- I think what was the most useful is patient self-report_ **## Discussion** In this 3-month pilot study, we applied anomaly detection to active and passive smartphone data in 15 subjects with schizophrenia. ==Of the 15 subjects, five experienced a clinical relapse, and of those three had sufficient data for anomaly detection.== For individuals who relapsed, the rate of anomalies detected in the passive data streams in the 2 weeks prior to relapse was 71% higher than the rate of anomalies detected in dates further away from relapse. Our results demonstrate how smartphone passive data hold potential to identify early warning signs for relapse in schizophrenia. (Based on 3 patient?) **- what about data for patient who did not relapse? how does those data look like?**