标签: 生命体征

Wetrainandvalidateasemi-supervised,multi-taskLSTMon57,675person-weeksofdatafromoff-the-shelfwearableheartratesensors,showinghighaccuracyatdetectingmultiplemedicalconditions,includingdiabetes(0.8451),highcholesterol(0.7441),highbloodpressure(0.8086),andsleepapnea(0.8298).Wecomparetwosemi-supervisedtrainingmethods,semi-supervisedsequencelearningandheuristicpretraining,andshowtheyoutperformhand-engineeredbiomarkersfromthemedicalliterature.WebelieveourworksuggestsanewapproachtopatientriskstratificationbasedoncardiovascularriskscoresderivedfrompopularwearablessuchasFitbit,AppleWatch,orAndroidWear.