Currently, artificial intelligence is used in the field of healthcare for administrative tasks, such as improving medical coding and in some cases of clinical utilize, such as improving reviews of diagnostic images of radiologists.
Here and there, however, some hospitals and healthcare systems start working on what some experts perceive as the next step in the evolution of AI in healthcare: forecasts.
In particular, the predictive analyzes in the field of AI-Exiblitive gains forthright attention by the directors of CIO and analytics.
One of the experts who perceives the forecast as the next limit of artificial intelligence in healthcare is Dr. Mintu Turakhia, a cardiologist at Veterans Affairs Palo Alto Healthcare System, Chief Medical and Scientific Officer at Irhytm and Professor of Medicine at the University of Stanford.
Turakhia has over 25 years of experience in caring for patient, research and research, data sciences and artificial intelligence, regulation of medical devices as well as the creation and commercialization of digital health products. In particular, he was a competition Apple Heart study.
He recently talked to Turakhia to immerse himself deeply in the AI forecast, talking about the steps needed for AI promotion from the current state to predictive possibilities, how predictive artificial intelligence can identify health conditions and enable preventive care, scaling of healthcare and creating better patients’ results with AI and integration of predictive artificial intelligence with IT systems of healthcare.
Q: You say that AI must go from the current state to predictive possibilities. What are the steps to get there?
AND. The most crucial early progress in artificial intelligence in healthcare were classification or recognition of patterns, starting with medical imaging. Deep learning algorithms are highly effective-in many cases they exceed clinicists-in identifying diagnoses on x-rays, ultrasound or electrocardiograms. AI can also stand out when measuring, such as estimating the left ventricle ejection fraction based on the ultrasound of the heart, which can often be burdensome and susceptible to human errors.
Recently, AI is used to obtain a diagnosis from electronic notes from health documentation, and even through conversational artificial intelligence with patients. It also falls into the classification.
However, the forecasts are different. He focuses on forecasting future results, and not on the identification of current countries. It is about using available data to estimate the risk of developing a disease or experience of a clinical event in the future.
For example, even if the outpatient outpatus outrient flicker does not detect the atria, the data it intercepts today, can still discover signals that predict the risk of AF down the line. Similarly, life symptoms, sleep patterns and activity data – often used for efficiency or sleep tracking – can instead be analyzed to predict the risk of hospitalization in the future.
It is about using life parameters, sleep data and activity so as not to earn “badges” on a smartwatch, but estimate the risk of developing hospitalization of heart failure.
Achieving predictive possibilities requires solid, generalized data sets related to clinical results. Historically, health data has been muted – imaging, ECG, smartwatch, medical documentation and hospitalization data of medical insurance exist independently. By combining these data sources at the patient’s level, you receive multidimensional and longitudinal data that can be used to develop AI models to predict the results in this data.
Another wave of artificial intelligence in healthcare will pass from diagnosing existing conditions for forecasting future health threats, paving the way to proactive and preventive care.
Q: You think that predictive artificial intelligence will be used more solidly to identify health conditions and allow preventive care. How
AND. Predictable artificial intelligence will allow us to identify the future risk of clinical diseases and events with greater precision. When I think about the development of artificial intelligence, I like to consider the 2×2 matrix: which is effortless or hard for people compared to what is effortless or hard for AI.
Example take an outpatient ECG monitoring again. The first step was to develop solid artificial intelligence to diagnose arrhythmia. We published it in 2019 (Hannun AW et al.) And since then hundreds of additional tests have shown this ability.
The next step is more intricate: the utilize of ECG to predict the future risk of atrial fibrillation. ECG can detect subtle structural and electrical changes in the heart that escalate the risk of AF. In combination with continuous ECG data – such as 14 days of monitoring – AI can identify critical patterns that people can miss. The integration of these patterns with a predictive risk model is calculated hard for people, but feasible and effortless to AI.
From there, predictive artificial intelligence can go further – estimating the risk of future results, such as stroke or heart failure, two conditions that are known to be caused by AF. Developing these predictive possibilities requires a combination of various data sets and conducting significant development works. However, potential benefits for early intervention and prevention are amazing.
Q: AI will generate better patient results, you suggest. How will it happen?
AND. Many people may not be aware of the remote monitoring of patient data began over 30 years ago. In the 1990s, manufacturers of implantable heart devices – such as starter starters and defibrillators – developed systems for remote monitoring of the function of the device and detection of arrhythmia.
Today, with progress in miniaturization, the sensor, what once required a visit to the office can be done at home – even at the smartwatch. For example, smartwatch algorithms can detect eternal irregular impulses and notify the user about the possibility of atrial fibrillation, enabling prior detection. This is already happening.
Looking to the future, the integration of many data streams – ECG, life parameters, sleep data and not only – in longitudinal models will allow AI to identify health threats before clinical events. For example:
- Anticipating the beginning of AF, heart failure or sleep apnea.
- Detection when a state similar to heart failure deteriorates, increasing the risk of hospitalization.
In these scenarios, clinicians, patients and healthcare systems can take proactive steps – such as confirmation of diagnosis, starting therapy or adaptation of drugs to reduce the risk of hospitalization.
Now, for this to work, and it must be solid in its performance. This means that his measures of accuracy and predictions – such as positive and negative predictive values - must be high. For example, if only 5% of all positive AI results are real, then 95% are false positives, which is not very useful and can even be harmful.
That is why AI will work best in fairly common conditions, because the identification of scarce diseases or events with high precision remains quite hard.
Q: You predict that hospitals and healthcare systems integrate predictive artificial intelligence with IT systems. How will they do it and for what purpose?
AND. There are two basic applications of predictive artificial intelligence in hospitals and healthcare systems.
First, at the patient level. This case of utilize is already underway. In outpatient care, clinicians often rely on basic risk results, which include a petite number of clinical variables. These results have restricted predictive accuracy.
AI improves these tools, integrating dozens and even hundreds of data points to generate more precise risk grades. Even if AI is not fully predictable, it can serve as a decision support, reducing improper differences in care. For example, AI can ensure that patients with atrial fibration receive anticoagulant treatment in accordance with clinical guidelines.
On the hospital side, several companies have developed early warning systems for sepsis, life -threatening complication of an overwhelming infection. Before the septic shock occurs, it is often too slow, and mortality rates reach 30-40%. Studies have shown that sepsis warning systems not only lead to better patients’ results, but also improve the compliance with clinicians of treatment protocols.
As a result, the quality of care can also improve.
Secondly, at the population level. In the case of integrated and based on the values of healthcare systems, predictive artificial intelligence can identify patients with the highest risk of using healthcare, usually visits to the ambulance and hospitalization. This allows the river to intervene to reduce pricey events.
Interestingly, the most effective solutions can be quite low-as home visits, regular telephone check-in, ensuring compliance with medicines or supporting family involvement.
Some healthcare systems even study generative “agents” or virtual nurses in order to carry out remote observations and monitor patients. The integration of predictive artificial intelligence with these tools has the potential to improve care, reduce costs and improve results.