Big data in the Cardiac ICU What can CNN can do for you? Kevin - PowerPoint PPT Presentation


 Big data in the Cardiac ICU 
 
 What can CNN can do for you? Kevin Maher, MD Professor of Pediatrics, Emory University School of Medicine Director, Cardiac Intensive Care, Children’s Healthcare of Atlanta Medical Director, Pediatric Technology Center Georgia Institute of Technology

• No Disclosures Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Overview • Clinical medicine and data in the CICU • CNN • Atlanta experience with CNN and CICU Data Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Critical Care Management/Decisions (+) Quality of decisions (-) Time

Hospital Course Discharge home health (+) Admission (-) illness Death

Medical Decisions • History and Physical • Laboratory data • Radiologic data • Anatomy • Physiology • Operation • Current course/clinical state • Medications • Genetic data • Study results (echo, etc) • MD Experience • MD Knowledge (book) Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Critical care data • About 150,000 waves of each type per day for each patient • Close to a million individual waves per patient, per day • In addition to laboratory, clinical diagnosis, radiologic, genetic, historical, and interventional data Children’s Healthcare of Atlanta, Emory University, Georgia Tech

ECMO support in Children Children’s Healthcare of Atlanta, Emory University, Georgia Tech

25-30 patients Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Critical Care Data • Too much data! • Most data is ignored • No way to visualize the data, understand or utilize • The subtleties of the waveforms that we simply don’t see Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Overview • Clinical medicine and data in the CICU • CNN • Atlanta experience with CNN and CICU Data Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Convolutional Neural Network • CNN is a deep learning algorithm that works well in image analysis • “Computer Vision” , how images are evaluated, categorized, identified • Does well with large data sets, and improves as data sets increase in size Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Consider a waveform as an image to be analyzed Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Convolutional Neural Networks “R” wave Convolution: extracting features from the image using filters/kernels Pooling reduces the dimensions of the data Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Overview • Clinical medicine and data in the CICU • CNN • Atlanta experience with CNN and CICU Data Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Methods: We conducted a binary classification for • differentiating whether a 30-sec ECG clips were from a ‘critical’ (post surgery) or ‘healthy’ (ready to transfer step- down) child. 51 children into training set • 10 most recent children into test set. • We obtained about 253k valid training • ECG clips and 74k test clips per each of the 3 leads. Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Methods A 64-layer 32 paths Convolution Neural Network (CNN) Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Methods • Train CNN on each of the 8 waveforms • Waveforms include I, II, III, ABP , RESP , Pleth, CVP , and SPO2 • Each waveform is segmented to 30 seconds • Discard segmentations with missing values • I: 261M valid data points => contain 2 billion numerical values • II: 327M valid data points • III: 214M valid data points Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Preliminary Results • Group 1 (ECG): CNN models separately trained on lead-I, II, III ECG waveforms, and a unified model combining all the three leads. • Group 2: ML models for heart rate variability (including LR, DT and RF) • Group 3: ML models (including LR, DT and RF) trained on vital signs, or lab results, or both. • Group 4 (combined): CNN model on ECGs combined with the ML models with vital signs and lab results Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Comparison of models

Results combined scores Post operative day The combined scores are computed by taking an average of all the predicted scores from 3-Lead waveforms, vital signs and lab results.

Dynamic changes of patients’ embeddings in a 2-D projection space Green dots: Red dots: Ready-to transfer Post-op to the floor critical Blue dots: A test patient’s embedding’s that change over time Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Post op day zero from Norwood operation, CICU Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Post op day 1 Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Post op day 2 Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Post op day 3 Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Post op day 4 Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Post op day 8 Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Day 9 post op Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Conclusions • Data, and knowledge exists in the critical care waveforms • Application of CNN to critical care data may enhance patient monitoring and patient management • The initial work presented is undergoing extensive clinical correlation and validation Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Conclusions • Additional goals include a real time “barometer” of clinical wellness • Opportunity for real time feed back for all medical decisions being made Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Research Team • Kevin Maher, MD CHOA/Emory • Alaa Aljiffry, MD CHOA/Emory • Yanbo Xu, PhD Georgia Tech • Jimeng Sun, PhD Georgia Tech • Siddharth Biswal, MS Georgia Tech • Shenda Hong, PhD Georgia Tech • IT Team CHOA/GT Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Innovation and Collaboration: 
 Advancing the Science and Treatment of Congenital Heart Disease Thank you

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

: A full trajectory of transfer prediction on a test patient who was transferred to step- down on Day 7. A threshold line: Median of the prediction scores across the full predictive trajectory Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Near Infrared Spectroscopy Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Despite anticoagulation, thrombosis persists Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

CNN results using ECG + BP arrest Inpatient, stable 2.Inpatient stable Discharge Critical 4. Critical Current accuracy is 97% to distinguish these 4 groups ate the patient groups Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Methods A 5-second ECG waveform contains 5 * 250Hz = 1,250 numeric values. A 1-hour ECG waveform contains 3600 * 250Hz = 900,000 numeric values. A 1-day ECG waveform contains up to 24 * 3600 * 250Hz = 21M numeric values. An ICU stay could range from 5 days up to 28 days (in our study cohort). Our goal: To develop a comprehensive computational algorithm that can efficiently learn patients’ dynamic physiological status from continuous ECG waveforms. Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Arterial wave forms Children’s Healthcare of Atlanta, Emory University, Georgia Tech

Methods (patient population) • Evaluate a cohort of patients with HLHS having undergone the Norwood operation • (newborn infants, high risk open heart surgical procedure) • Patients arriving from the operating room critically ill • The same cohort, when being transferred to the general cardiology floor considered “healthy” • Evaluate continuous waveform data, lab, EMR, & vital signs Children’s Healthcare of Atlanta, Emory University, Georgia Tech