HARSHA URLAM & SWETA SINGH at Cypher 2017
summary
Predicts diabetes related distress (not clinical depression) from mobile sensing collected by the StayClose app; starts with ~20,000 rows and 215 noisy features, reduces them via factor analysis to 215 metrics → 24 factors, and uses imputation (median/mean) before training classifiers. Random Forest and XGBoost perform best; labeled sample size reported at ~5,549 records and final model reports about 86% accuracy with <12% type‑II error. Pipeline includes an app alert to clinicians and notes missing clinical variables as a limitation.
AI-written summary · about
chapters (6)
- 0:00Project introduction
- 2:05Understanding the problem
- 4:32Methodology and goals
- 7:15Data collection and cleaning
- 14:34Model building and evaluation
- 21:09Results and future outlook
related talks
06mPredicting Subjective Measures of Social Anxiety from Sparsely Collected Mobile Sensor Data
40mBringing mental health data to doctors
33mData Science in Healthcare: Beyond the Hype
17mTransforming Medical Records Into a Heart Attack Prediction Model
39mBridging health inequalities through machine learning
50mDiabetes: Calculating Insulin by Automated Carb Counting Using AI, ML
20mML applications in healthcare - Saudi Arabian hospital case study
20mIdentification of Imminent Suicide Risk Among Young Adults using Text Messages