AI to help doctors develop personalized treatments

No patient is the same. Sounds obvious, yet often doctors treat patients with the same diagnosis in a similar way. Our team has turned to AI — and have developed a software, described in our latest papers, to help doctors personalize treatments for different patients with the exact same diagnosis.

We’ve combined information from clinical treatment guidelines with historical electronic health records data. The result was a surprise: we’ve found that more often than not, there are multiple treatment options, some of which can go unconsidered in the treatment planning process. Our model could enable doctors to consider the best-practice experiences of their colleagues to help them make more informed decisions about a treatment for a specific patient. ^{1 2}

Typically, doctors use evidence-based medicine that integrates available research with clinical expertise and patient data to make the most informed treatment decisions. This means mainly relying on the results of randomized controlled trials. But such trials don’t always cover all clinical conditions and usually involve a relatively small number of study subjects. They also attempt to remove as many variables as possible to remain random — meaning that the subjects under study can often represent a relatively homogeneous group that’s not very representative of the real world.

To create the software, our AI first combed through the database of Atrius Health — a large health care provider in eastern Massachusetts with more than 20 years of data for approximately 2.5 million patients. We specifically analyzed the records of patients suffering from one of three chronic diseases: hypertension, type 2 diabetes, and hyperlipidemia or high cholesterol. We then took the reference clinical treatment guidelines for each disease and manually extracted the set of recommended drugs and the factors that determine which patients should be prescribed which drug.

Next, we took historical patient data from the Atrius electronic health records and identified the so-called decision points: doctor visits where a patient’s disease status was not controlled. For each decision point, we extracted all available information about the patient leading up to that time, such as any drug treatment decisions, lab test results and other diseases, and whether the disease was under control during follow-up visits.

Armed with all this data, we trained disease-specific machine learning models to identify similar decision points elsewhere in the data. The result was a so-called “precision cohort” — a group of patients that closely match based on their individual characteristics and their precise clinical situation.

These cohorts have allowed us to study the responses of patients to various treatments by different doctors in the same organization under similar clinical scenarios. This way, we’ve been able to determine the best-outcome treatment options – those that produce the highest level of disease control — in the organization for patients similar to the individual patient in similar clinical situations.

For instance, a diagram below shows an example of observed treatments and associated outcomes for a precision cohort. The cohort consists of the data of patients all similar to a specific individual patient under the same clinical situation. In the example, all patients in the cohort have a diagnosis of hypertension and are treated with an angiotensin receptor blocker (ARB) only. In the middle column, we’ve detailed the treatment options of various doctors in the practice, and the width of the “prong” shows the relative size of the cohort choosing that option.

Precision cohort visualization in the Electronic Health Records

We’ve noticed that the most common choice was to make no change in the ARB drug class. Under the no-change scenario, only 40 percent of those patients had controlled blood pressure at the follow-up measurement. The prongs above the no-change group all had an increase in percent controlled on follow-up; the treatment cohorts in green had a statistically significant change in percent controlled. The prongs below the no-change group had a lower percent controlled on follow-up, and red prongs indicate a statistically significant change.

Better options available

This example clearly shows that even for a relatively homogeneous cohort, doctors in the same organization selected many different treatment options. Some treatments were selected more often than others and some treatments had better associated disease control outcomes than others.

We’ve learned that in the vast majority of cases across the three diseases, there were multiple other treatment plans than the one a specific doctor had picked. For hypertension, 65.1 percent of cases could have presented doctors with multiple treatment paths based on an analysis of precision cohorts, for Type 2 diabetes – 37.7 percent, and for high cholesterol – 75.3 percent [2]. If applied in a clinical setting, surfacing such data could enable clinicians with accurate, personalized information that can help them to make tailored and data-driven treatment decisions for their patients.

Our method isn’t limited to these three diseases and can be applied to any disease-treatment situation, where relevant past experiences of patients in similar situations can help better inform clinical decision making.

We think that our software could help physicians to be aware of the best practice experiences of other doctors and the lessons they’ve learned from the myriads of doctor-patient interactions captured in electronic health records. That, combined with traditional treatment guidelines, could potentially help them make more informed decisions about the best treatment for individual patients.

However, our research isn’t over yet. One of the biggest challenges is the need for large enough observational datasets to cover the variety of diseases and treatment options. We are now exploring several approaches to address this challenge – including more efficient ways to use the existing data, such as weighting data elements instead of filtering some out and combining multiple datasets.

With digital archives and electronic records, there is no lack of raw data. But inundating doctors with overwhelming amounts of data is not useful. Our model is a tool to extract and visualize the insights in the data – that could help them make more informed decisions.