GHC 14 A Statistical Clinical Decision Support Tool

#GHC 14 A Statistical Clinical Decision Support Tool in Telemonitoring Using Predictive Analytics © 2014 Celeste Fralick, Ph. D. Principal Engineer, Chief Data Scientist Analytic Products & Technologies Internet of Things Group Intel Corporation 2014

Agenda § § § Challenge and Approach Segregating Clinical & Statistical Thresholds Classification Model Predicting Future Episodes Summary & Future Studies 2014

Challenge & Approach Broadband & Private Cloud Clinician derived thresholds colors Elder with COPD or CHF Clinicians monitoring vitals • Provide clinician with robust statistical thresholds using Statistical Process Control (SPC) & Western Electric Rules • Patient’s own data will drive thresholds; normal variation seen • Use predictive analytics to forecast the probable classification • Predict next day’s vital sign reading and whether it’s statistically normal for that patient • Provide opportunity for clinicians to intervene earlier, lower costs & hospital readmissions, and improve patient outcomes.

Forced Expiratory Volume in 1 sec in L (FEV 1, P#18) Peak Flow with Two Thresholds (Clinician & Statistical) FN FN Clinician Derived Threshold FP FP TP Date in Time Series (7/09 -7/10) Point Detection Location One point Detects a shift in the mean, an 1 beyond increase in the standard deviation, or Zone A a single aberration in the process. Nine points in a row in a single 2 (upper or Detects a shift in the process mean. lower) side of Zone C or beyond Rule Confusion matrix & ROC available Intervention Test Hyp State Statistically Derived State Clinician Derived Combined State Result Intervened TP H 0 1 Threshold violation 1, 1 4 Did not intervene but should have (statistically) FN H 1 1 Threshold violation 0 No threshold violation 1, 0 4 Intervened but shouldn't have (statistically) FP H 1 0 No threshold violation 1 Threshold violation 0, 1 4 No intervention necessary TN H 0 0 No threshold violation 0, 0 47

(Polynomial Shift) Parameter Result Disease Severity Severe Mean / N 0. 884 / 59 UCL / LCL 1. 343 / 0. 425 Std Dev 0. 883525 Variance 0. 83 Skewness / Kurtosis 0. 03458/ 0. 593806 Distribution Normal 2 Mixture Training Gen R Square 0. 875022 Validation Gen R Square 0. 938169 Date in Time Series (7/09 -7/10) Used Random K Fold (5 folds) with 6 activation nodes and squared penalty. No boosting model. Misclassification rate: 0. 083

σ edge misclassifications Deeper analytics provide deeper insight: • Pre- & post predictive algorithm of SPC zones generally agree • Severe & Very Severe COPD patients provide better predictive models • Mean FEV 1 FP→FN % negligible increase for moderate COPD

Disease Weight Blood pressure FEV 1 PEF Sp. O 2 Pulse CHF Increases Decreases Increases/ arrhythmia COPD Increases in severe cases Increases Decreases Increases/ arrhythmia

Summary § Provided Clinical Decision Support using elder’s own vitals rather than generalizing § By applying a unique neural net algorithm, able to classify successfully § Personalized prediction of vital sign may impact disease outcome, re-hospitalization 2014

Future Studies § Additional vital signs for specific disease § Additive algorithms to increase fit and accuracy § Reduce sources of errors via tightly controlled clinical study § Consider better neural net penalty to address overfitting 4/16/2013 crf 2014 9

Got Feedback? Rate and Review the session using the GHC Mobile App To download visit www. gracehopper. org 2014

Backup 2014

Random K-Fold • Investigated a random K-fold (5) for predicting values & classification – Cross-validation technique for small sample sizes in neural networks Responses xval Input TR Data TR TR 4/16/2013 VAL Nodes 5 x xval TR X -bar xval VAL Xbar xval crf Act Func TP Transformation TR FN FP TN 12