Today, artificial intelligence is being used across healthcare for administrative tasks, such as improving medical coding, and for certain clinical use cases, such as enhancing radiologists’ reviews of diagnostic imagery.
Here and there, though, some hospitals and health systems are beginning work on what some experts view as the next step in the evolution of AI in healthcare: prediction.
AI-enabled predictive analytics, in particular, is getting a fair deal of attention by CIOs and analytics executives.
One expert who sees prediction as the next frontier for AI in healthcare is Dr. Mintu Turakhia, a cardiologist at the Veterans Affairs Palo Alto Healthcare System, chief medical and scientific officer at vendor iRhythm and professor of medicine at Stanford University.
Turakhia has more than 25 years of experience in patient care, outcomes research and trials, data science and artificial intelligence, medical device regulation, and the creation and commercialization of digital health products. Notably, he was the co-principal investigator of the landmark Apple Heart Study.
Healthcare IT News spoke recently with Turakhia to dive deep into AI prediction, talking about the steps needed to advance AI from its current state to predictive capabilities, how predictive AI can identify health conditions and enable preventive care, scaling healthcare and creating better patient outcomes with AI, and integrating predictive AI into healthcare information systems.
Q. You say AI needs to advance from its current state to predictive capabilities. What are the steps to getting there?
A. The most significant early strides in artificial intelligence in healthcare have been in classification, or pattern recognition, starting with medical imaging. Deep learning algorithms are highly effective – in many cases outperforming clinicians – at identifying diagnoses on X-rays, ultrasounds or electrocardiograms. AI also can excel at measurements, such as estimating left ventricular ejection fraction from cardiac ultrasounds, which often can be cumbersome and prone to human error.
More recently, AI is being applied to derive diagnoses from electronic health record notes and even through conversational AI with patients. This too falls into classification.
Prediction, however, is different. It focuses on forecasting future outcomes rather than identifying current states. It’s about using available data to estimate the risk of developing disease or experiencing a clinical event in the future.
For example, even if an ambulatory ECG monitor does not detect atrial fibrillation today, the data it captures still can uncover signals that predict AF risk down the line. Similarly, vital signs, sleep patterns and activity data – often used for fitness or sleep tracking – can instead be analyzed to predict the risk of future heart failure hospitalizations.
It’s about using vital signs, sleep and activity data not to earn fitness or sleep “badges” on your smartwatch but to estimate the risk of developing a heart failure hospitalization.
Reaching predictive capabilities requires robust, generalizable datasets connected to clinical outcomes. Historically, health data has been siloed – imaging, ECG, smartwatch data, medical records and medical insurance hospitalization data all exist independently. By linking these data sources at the patient level, you get multidimensional and longitudinal data that can be used to develop AI models to predict the outcomes in that data.
The next wave of AI in healthcare will shift from diagnosing existing conditions to forecasting future health risks, paving the way for proactive and preventive care.
Q. You believe predictive AI will be used more robustly to identify health conditions and enable preventive care. How so?
A. Predictive AI will enable us to identify future risks of health conditions and clinical events with greater precision. When I think about AI development, I like to consider a 2×2 matrix: what is easy or hard for humans versus what is easy or hard for AI.
Take ambulatory ECG monitoring again as an example. The first step was to develop robust AI for diagnosing arrhythmias. We published this in Nature Medicine in 2019 (Hannun AW et al.), and since then, there have been hundreds of additional studies demonstrating this capability.
The next step is more complex: using the ECG to predict future risk of atrial fibrillation. The ECG can detect subtle structural and electrical changes in the heart that increase the risk of AF. When combined with continuous ECG data – such as 14 days of monitoring – AI can identify critical patterns humans might miss. Integrating these patterns into a predictive risk model is computationally difficult for humans but feasible and easy for AI.
From there, predictive AI can go further – estimating the risk of future outcomes like stroke or heart failure, two conditions known to be caused by AF. Developing these predictive capabilities requires linking diverse datasets and conducting significant development work. However, the potential benefits for early intervention and prevention are extraordinary.
Q. AI will generate better patient outcomes, you suggest. How is this going to happen?
A. Many people may not realize remote monitoring of patient data began more than 30 years ago. In the 1990s, manufacturers of implantable cardiac devices – such as pacemakers and defibrillators – developed systems to remotely monitor device function and detect arrhythmias.
Today, with advances in sensor miniaturization, what once required an office visit now can be done at home – even on a smartwatch. For example, smartwatch algorithms can detect sustained irregular pulses and alert the user to the possibility of atrial fibrillation, enabling earlier detection. This already is happening.
Looking ahead, the integration of multiple data streams – ECGs, vital signs, sleep data and more – into longitudinal models will enable AI to identify health risks before clinical events occur. For example:
- Predicting the onset of AF, heart failure or sleep apnea.
- Detecting when a condition like heart failure worsens, increasing the risk of hospitalization.
In these scenarios, clinicians, patients and health systems can take proactive steps – such as confirming diagnoses, starting therapies or adjusting medications to reduce hospitalization risk.
Now, for this to work, AI needs to be robust in its performance. This means its measures of accuracy and prediction – such as positive and negative predictive values – must be high. For example, if only 5% of all positive AI results are true, then 95% are false positives, which is not very useful and could even be harmful.
This is why AI will work best for fairly common conditions, as identifying rare diseases or events with high precision remains quite challenging.
Q. You predict hospitals and health systems will be integrating predictive AI into information systems. How will they do so, and to what end?
A. There are two primary applications of predictive AI within hospitals and health systems.
First, at the patient level. This use case already is underway. In outpatient care, clinicians often rely on basic risk scores that consider a small number of clinical variables. These scores have limited predictive accuracy.
AI enhances these tools by integrating dozens or even hundreds of data points to generate more precise risk assessments. Even when AI isn’t fully predictive, it can serve as decision support, reducing inappropriate variations in care. For example, AI can ensure patients with atrial fibrillation receive anticoagulation therapy as recommended by clinical guidelines.
On the inpatient side, several companies have developed early warning systems for sepsis, a life-threatening complication of overwhelming infection. By the time septic shock occurs, it often is too late, with mortality rates reaching 30-40%. Studies have shown that sepsis alert systems not only lead to better patient outcomes but also improve clinician adherence to treatment protocols.
As a result, the quality of care can improve as well.
Second, at the population level. For integrated and value-based health systems, predictive AI can identify patients at the highest risk for healthcare utilization, typically emergency room visits and hospitalizations. This enables upstream interventions to reduce costly events.
Interestingly, the most effective solutions can be quite low-tech – such as home visits, regular check-ins by phone, ensuring medication adherence or supporting family engagement.
Some health systems are even exploring generative AI “agents” or virtual nurses to conduct remote follow-ups and patient monitoring. The integration of predictive AI with these tools holds the potential to enhance care, reduce costs and improve outcomes.
Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication
