BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT DTSTART:20250309T030000 TZOFFSETFROM:-0800 TZOFFSETTO:-0700 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20251102T010000 TZOFFSETFROM:-0700 TZOFFSETTO:-0800 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20250904T033743Z UID:35793245-CE90-479F-B768-0FA02BA64EE5 DTSTART;TZID=America/Los_Angeles:20250827T180000 DTEND;TZID=America/Los_Angeles:20250827T190000 DESCRIPTION:Wearable health devices are widely adopted tools that continuou sly collect physiological data\, and thus\nmake possible passive\, continu ous\, data-driven assessments of people’s health. Algorithms can detect\ nacute illnesses\, like COVID-19\, by distinguishing physiological time se ries data that contains anomalous\npatterns\, often during sleep\, from pr esumably healthy\, stable baseline data. However\, some individuals’\nba seline states contain shifts and fluctuations that can look like anomalous patterns but are actually\ndynamic characteristics of that individual’s baseline. Precision dropped substantially (-23.4% in AUC) for\nindividual s with dynamic baselines in COVID-19 detection experiments because anomaly detection\nalgorithms are designed to rely on the stability of baseline s tates. We used 5 million nights of sleep data\nto investigate new approach es to modeling dynamic baselines and show our temporal model improves\nsep arability by 4-10x across acute health conditions (COVID-19\, flu\, and fe ver). With this model\, we\ndrastically recovered performance (+19.4% in A UC) with large reduction in false positive errors.\nModeling how people ar e dynamic over time is essential not only to identifying anomalous health states\nbut also to building robust health monitoring systems in the real world\, where people are inherently\ndynamic\, and empowering individuals to take data-informed actions that meaningfully preserve their\nhealth.\n\ nSpeaker(s): Varun\n\nVirtual: https://events.vtools.ieee.org/m/497194 LOCATION:Virtual: https://events.vtools.ieee.org/m/497194 ORGANIZER:sudip453@gmail.com SEQUENCE:18 SUMMARY:Five Million Nights: Temporal Representations in Anomaly Detection Algorithms for Wearable Health URL;VALUE=URI:https://events.vtools.ieee.org/m/497194 X-ALT-DESC:Description: <br /><p>Wearable health devices are widely adopted tools that continuously collect physiological data\, and thus<br>make pos sible passive\, continuous\, data-driven assessments of people&rsquo\;s he alth. Algorithms can detect<br>acute illnesses\, like COVID-19\, by distin guishing physiological time series data that contains anomalous<br>pattern s\, often during sleep\, from presumably healthy\, stable baseline data. H owever\, some individuals&rsquo\;<br>baseline states contain shifts and fl uctuations that can look like anomalous patterns but are actually<br>dynam ic characteristics of that individual&rsquo\;s baseline. Precision dropped substantially (-23.4% in AUC) for<br>individuals with dynamic baselines i n COVID-19 detection experiments because anomaly detection<br>algorithms a re designed to rely on the stability of baseline states. We used 5 million nights of sleep data<br>to investigate new approaches to modeling dynamic baselines and show our temporal model improves<br>separability by 4-10x a cross acute health conditions (COVID-19\, flu\, and fever). With this mode l\, we<br>drastically recovered performance (+19.4% in AUC) with large red uction in false positive errors.<br>Modeling how people are dynamic over t ime is essential not only to identifying anomalous health states<br>but al so to building robust health monitoring systems in the real world\, where people are inherently<br>dynamic\, and empowering individuals to take data -informed actions that meaningfully preserve their<br>health.</p> END:VEVENT END:VCALENDAR