Using routine data to inform routine clinical problems Allan J Walkey, MD, MSc Associate Professor of Medicine The Pulmonary Center Evans Center of Implementation and Improvement Sciences
ICU Medicine Poorly- Well- Well-managed characterized Characterized clinical problems Clinical Problems Clinical Problems
Intersection of well-characterized problems = Intersection of well-characterized problems = poorly characterized problem poorly characterized problem New-Onset Atrial Fibrillation (AF) during Pneumonia and Sepsis What does it mean and what should I do when patients with sepsis develop AF?
My First Forays into “Big Data” Good data: Data that exists! AHRQ State Inpatient Databases include 240 data elements from every hospitalization California = 3.9 million hospitalizations/year ~1 billion data points
Characterizing aClinical Problem: Epidemiology New-onset AF in Sepsis Is risk of new-onset AF increased in sepsis? YES! By nearly 7-fold; 10% of patients Is it bad? YES! Increased risk of death and stroke. How long so those risks last? A long time! Walkey AJ et al. Chest. 2014;146(5):1187-1195. Walkey AJ et al. JAMA 2011; 306 (20) 1615
CharacterizingProblems: Better data Could we learn more about AF if we could accurately detect AF in banked ECG waveforms within EHR data? What if we could predict AF onset…could we intervene beforehand to prevent AF? With what?
CharacterizingProblems: Better data Database: MIMIC III Open Source ICU data ~ 2000 sepsis patients with electronic health records linked to banked waveform data
More Data, More Problems Variable frequency complex demodulation Noise! (VFCDM) based time-frequency spectral (TFS) method to detect and remove noisy segments Bashar S, Walkey AJ, Mcmanus DD, Chon K. IEEE Access 2019; 7: 88357 - 88368
AF Feature Detection QRS Complexes P-Waves Hossain B, Bashar S, Walkey AJ, McManus, DD Chon K. IEEE Access 2019: 128869
AF Prediction WIP
Better-managed clinical problems Question: What is most effective treatment to reduce heart rate during AF in sepsis? phenylephrine Need: Really granular data Philips eICU open source: norepinephrine 200,000 admissions,100 ICUs Vital signs resolution: 1 min.
Other Ongoing Studies • Predicting cardiovascular events after sepsis using EMR data to better target treatments • Use of novel continuous regression discontinuity designs to evaluate implementation effectiveness • Automated extraction of EMR data for outcome ascertainment in pragmatic trials of AF treatments
Summary • So many poorly defined clinical problems + so much underused data = learning opportunity • Diverse methodological + subject expertise = necessity in data science • AF + Sepsis = Bad… but maybe we can change that using routinely collected data to enable a Learning Healthcare System
Acknowledgements Funding/Support R01HL136660 R01HL139751 K01 HL116768 Boston University School of Medicine Department of Medicine Career Investment Award
Recommend
More recommend
