AI, NLG, and Machine Learning

AI Could Predict Your Lifespan

Researchers have developed artificial intelligence tools that may help to identify health problems faster and more efficiently.

By Ingrid Fadelli
January 22, 2020

Artificial intelligence (AI) has taken major steps forward over the past few years, with researchers worldwide developing a variety of tools that make predictions by analyzing large quantities of data.

These tools could have important implications for a variety of fields, including transportation, manufacturing, and healthcare.

Perhaps one of the most interesting ways in which AI could be applied in real-life settings is by helping healthcare professionals to diagnose patients and to predict the risk of developing particular health conditions.

A team of researchers at Geisinger Medical Center in Danville, Pennsylvania, has been conducting a series of studies exploring the potential of AI for the detection of specific heart conditions, as well as for predicting people’s overall health and life expectancy.

Researching how AI could be applied in healthcare settings

Studies investigating the possible use of AI and machine learning (ML) techniques in healthcare have become increasingly popular.

These techniques can analyze vast quantities of data quickly and effectively, attaining results that could never be achieved using more conventional computational tools.

They could thus open up new exciting possibilities for the early detection of physiological abnormalities or diseases, particularly those that can be identified based on brain recordings or images.

Geisinger has been conducting research exploring the potential of AI in healthcare for several years, and they have recently attained very promising results.

Geisinger's role

“Our institution, Geisinger, is unique with regard to healthcare data resources because it was an early adopter of Epic electronic health records (EHR), as well as a centralized data warehouse, and we are an integrated health system serving a very stable patient population, meaning that we have detailed, longitudinal data available for many patients,” Christopher Haggerty and Brandon Fornwalt, the two main researchers behind Geisinger’s recent studies, told

“In our recent work, we thus chose to take a unique approach to the application of machine learning by asking if we could predict relevant future events, such as onset of atrial fibrillation or all-cause mortality, from a widely used medical test, the 12-lead electrocardiogram (ECG).”

ai lifespan
Summary of model performance to predict one-year mortality using ECG data.

Predicting future health, based on ECG data

Geisinger has collected patient health records for almost 30 years, and it has carried out millions of ECG studies. In addition, being a medical center, it also possesses accounts of the future health of the patients that the ECG belonged to, such as whether or not they developed a particular condition or died in the years after the ECG was taken.

The value of data

This large amount of data might now prove incredibly precious, as it could be used to train state-of-the-art machine learning algorithms, linking particular features in an ECG to patients’ future health outcomes.

To start testing this exciting possibility, Haggerty, Fornwalt, and their colleagues first labeled each of the ECG recordings collected by Geisinger with information related to the patient’s outcome.

Subsequently, they used this labeled data to train a deep Convolutional Neural Network (CNN), hoping that it would ultimately learn to predict patient health outcomes based on their ECG scans.

ai lifespan
Neural network architecture for mortality prediction from echocardiography videos and electronic health record data.

“The only inputs we fed to our model were the voltage data measured during the ECG,” Haggerty and Fornwalt said.

“We used standard cross-fold validation techniques, meaning that we trained and tested the model on distinct patient data, such that the model was not tested on data it had seen before.”

Continuing to test

The researchers at Geisinger have already carried out a substantial number of studies testing the effectiveness of AI in predicting specific health outcomes. Some of these studies have been published in reputable scientific journals, while others are still being reviewed and are set to be published soon.

A large portion of their research has focused on predicting mortality, regardless of a patient’s cause of death, mainly because it is a “clean” label that is well documented in their databases.

“Our different studies have focused on different patient populations (congenital heart disease, heart failure, etc.) or different forms of input data (for example, using structured, clinically reported measurements vs. raw echocardiographic or electrocardiographic images/signals),” Haggerty and Fornwalt said.

“We are also working toward a broader array of clinical end points to predict, but that requires a tremendous amount of work to curate and clearly define these labels from the EHR data.”

The future of predictive AI

The studies conducted by Haggerty and Fornwalt have already achieved promising results, particularly in the prediction of death and heart irregularities, such as arrhythmia. Haggerty and Fornwalt are now investigating the potential of machine learning techniques for the prediction of other clinically relevant events.

“Our ECG model work is among the first to demonstrate the ability to predict future clinical events from ECG data, so we are very excited to explore the potential for this framework more broadly,” Haggerty and Fornwalt said.

“In other words, what other clinically relevant events can we predict using this approach to provide clinically relevant information to physicians and health systems to improve patient outcomes?”

Future plans and possibilities

Overall, Haggerty, Fornwalt, and their colleagues at Geisinger have found that machine learning models can predict health outcomes with a high level of accuracy, even in cases where a physician would interpret a patient’s ECG as “normal.”

This could have enormous implications for the future of healthcare, as it suggests that the computer models are identifying features predictive of mortality within ECGs that physicians do not currently recognize as abnormal and could thus completely alter the way in which we interpret ECGs.

Other healthcare applications

The researchers believe that AI could have a number of interesting uses in healthcare settings, for instance by assisting physicians caring for individual patients (e.g., a radiologist reading a study with AI assistance).

“We are also thinking about opportunities for machine learning to help health systems caring for a population of patients,” Haggerty and Fornwalt said. “That is, using the output of machine learning to help care teams prioritize where and how to deploy limited resources to maximize the overall health benefit to the population. This is a very challenging task.”

Despite the huge potential shown by some of the tools developed by this team of researchers, several limitations still need to be overcome before we can witness the widespread use of machine learning in healthcare.

ai lifespan
Accuracy for three cardiologists to correctly identify the true positive ECG (dead within a year) when presented with two 'normal' ECGs corresponding to a paired set of a true positive and true negative.

However, with the vast amount of medical data available today, as AI improves, it will most likely take on a central role in healthcare, helping physicians to review and synthesize the relevant data for a given patient, as well as to compare their patient to the patterns and associated outcomes from past patients, providing valuable insight to make more effective diagnostic and treatment decisions.

“One piece that will be critical in advancing the field is evaluation in a prospective setting,” Haggerty and Fornwalt said.

“That is, taking a rigorously validated model and using it in a controlled research study (a randomized clinical trial approach, for example) to determine if the model predictions were accurate in a new patient sample, and most importantly, if the predictions were able to positively influence patient care in some way. We see that evaluation as the ultimate goal for the work we are doing at Geisinger.”

Continued research & experimentation

Haggerty, Fornwalt, and their colleagues at Geisinger plan to continue their research into how AI could be used to predict patients’ health outcomes, focusing on several different health conditions. Ultimately, they hope to start testing the models they developed in clinical settings, to further evaluate their effectiveness.

“Generally, our goals for continuing the work described here are to begin clinically deploying and testing these initial models (e.g., predicting incident atrial fibrillation with ECG) within our health system to quantify the impact and to continue developing more complex models (combining data from multiple input sources) to improve accuracy, as well as defining and validating new end points to expand the clinical applicability,” Haggerty and Fornwalt said.