Skip to main content

We’ve updated our Terms & Conditions and Privacy Policy. By using this site, you agree to these terms.

Machine learning studies

predicting-survival

Predicting survival from large echocardiography and electronic health record datasets

Predicting patient outcomes (e.g., survival) following echocardiography is primarily based on ejection fraction (EF) and comorbidities. However, there may be significant predictive information within additional echocardiography-derived measurements combined with clinical electronic health record data.

Machine learning models achieved significantly higher prediction accuracy (all AUC >0.82) over common clinical risk scores (AUC = 0.61 – 0.79), with the nonlinear random forest models outperforming logistic regression (p <0.01). The random forest model including all echocardiographic measurements yielded the highest prediction accuracy (p < 0.01 across all models and survival durations). Only 10 variables were needed to achieve 96% of the maximum prediction accuracy, with 6 of these variables being derived from echocardiography. Tricuspid regurgitation velocity was more predictive of survival than LVEF. In a subset of studies with complete data for the top 10 variables, multivariate imputation by chained equations yielded slightly reduced predictive accuracies (difference in AUC of 0.003) compared with the original data.

 
deep-neural-network

A deep neural network to enhance clinical survival predictions based on heart imaging

We have demonstrated the potential predictive abilities of structured, clinical measurements derived from heart imaging (echocardiograms), but these measures do not capture all the data and insight that can be derived from rich imaging data. Hence the application of deep neural networks may add new insight.

Here we show that a large dataset of 723,754 clinically-acquired echocardiographic videos (~45 million images) linked to longitudinal follow-up data in 27,028 patients can be used to train a deep neural network to predict 1-year survival with good accuracy. We also demonstrate that prediction accuracy can be further improved by adding highly predictive clinical variables from the electronic health record. Finally, in a blinded, independent test set, the trained neural network was more accurate in discriminating 1-year survival outcomes than two expert cardiologists. These results therefore highlight the potential of neural networks to add new predictive power to clinical image interpretations. Download the paper

deriving-clinically-relevant-insights

Deriving clinically relevant insights from multi-modal interpretable neural networks

For many applications in medicine, advances in neural networks are unlocking new capabilities in predictive modeling (e.g., in medical image analysis). However, neural networks generally suffer from limited interpretability, which is an important consideration in medical applications. We are therefore interested in exploring ways to enhance the interpretability of neural networks to address this limitation. As proof of concept, we have developed and demonstrated the value of using an interpretable neural network for predicting an important clinical outcome (1-year mortality) using multi-modal clinical data including electronic health record (EHR) data and 31,278 echocardiographic videos from 26,793 patients.  We generated separate models for EHR data, numeric variables derived from videos and pixel data from raw echocardiographic videos. The interpretable multi-modal model maintained performance compared to non-interpretable models and performed significantly better than a model using a single modality. This allows us to derive clinically relevant insights from variable importance ranking. Download the paper

data-driven-pop-health

Toward data-driven population health management of heart failure

One application for the types of machine learning models we are developing and studying is risk stratification of specific groups of patients for effective population health management, such as patients with heart failure (HF). Machine learning provides a transformational approach to predicting HF admissions and mortality, and may reveal unique insights into pathophysiology and management options through identification of the most important  predictors. By providing these data-driven risk insights, and incorporating potential interventions, we believe we can improve patient outcomes on a large scale. 

As a proof-of-concept, we identified 11,327 patients with HF who had previously undergone 44,656 echocardiograms. Linear logistic regression and non-linear random forest models were developed to predict 1-year mortality and 6-month hospitalization after echocardiography. Both machine learning models achieved an AUC>0.75, with random forest  performing slightly better than logistic regression (mortality: 0.79±0.01; hospitalization: 0.79±0.02). Hemoglobin, lymphocytes, NT-proBNP and creatinine were the four most important variables for predicting both adverse outcomes, which provide important pathophysiologic insights. 

This work has formed the basis of one of the first randomized and matched interventional clinical trials based on insights from machine learning with the goal of improving patient outcomes.

Clinical trial information

AHA Journal paper

 
advanced-machine-learning-in-action

Advanced machine learning in action: identification of intracranial hemorrhage on computed tomography scans of the head with clinical workflow integration

A deep convolutional neural network was trained on 37,074 studies and subsequently evaluated on 9,499 unseen studies (~2 million 2D images). The predictive model was implemented prospectively for 3 months to re-prioritize “routine” head CT studies as “stat” on real-time radiology worklists if an ICH was detected. Time to diagnosis was compared between the re-prioritized “stat” and “routine” studies. A neuroradiologist blinded to the study reviewed false positive studies to determine whether the dictating radiologist overlooked ICH. The model achieved an area under the ROC curve of 0.85 (0.84–0.86). During implementation, 94 of 347 “routine” studies were re-prioritized to “stat”, and 60/94 had ICH identified by the radiologist. Five new cases of ICH were identified, and median time to diagnosis was significantly reduced (p < 0.0001) from 512 to 19 min. In particular, one outpatient with vague symptoms on anti-coagulation was found to have an ICH which was treated promptly with reversal of anticoagulation, resulting in a good clinical outcome.

Download the paper