Crunching Medicare Data Could Be The Key To Predicting Parkinson’s Disease — And Many Other Diseases Could Be Next

People with Parkinson’s disease, a debilitating movement disorder, exhibit symptoms like tremors and limb stiffness that get progressively more drastic as the condition worsens. But before a patient’s symptoms become pronounced, doctors have no reliable way to recognize Parkinson’s early on, or identify that a patient is on track to develop the disease.

A new approach to healthcare data analysis may change that: After crunching the Medicare claims data of more than 200,000 people, researchers from the Washington University School of Medicine in St. Louis developed an algorithm to predict if a patient will one day be diagnosed with Parkinson’s, or not.

The algorithm – which achieved 73-83% accuracy in study results – could ultimately help doctors reduce unnecessary testing and diagnose Parkinson’s cases faster and earlier. With further refinement, the work may become a standard-bearer for utilizing administrative healthcare data in clinically useful ways.

Algorithms and artificial intelligence (AI) are infiltrating the healthcare industry in a variety of applications – many of which relate to disease detection and prevention.

Researchers in Italy, for example, just developed an AI-powered algorithm able to detect Alzheimer’s in brain scans with 86% effectiveness. Startups in the AI medical space include Freenome (which uses machine learning to assess biological signals in the blood), Recursion Pharmaceuticals (which scrutinizes cell images to find new uses for existing drugs), and plenty of others.

But in each of the examples above, the AI assesses or relies on clinical content. The Washington University researchers, on the other hand, created a predictive model to identify Parkinson’s disease using only the demographic information and diagnosis & procedure codes from administrative claims.

The researchers launched the study — which was partly funded by the National Institutes of Health (NIH), Michael J. Fox Foundation, and American Parkinson Disease Association — knowing that almost all Parkinson’s diagnoses are preceded by years of mild but progressively worsening symptoms. For the purposes of their study, de-identified Medicare records provided a data set for investigating how the medical histories of people who ultimately develop Parkinson’s differ from those of people who never develop the disease.

To develop their algorithm, the WU team first analyzed five years of claims histories of ~208,000 Medicare patients aged 66-90. About 43% of patients in the data set (~90,000) were diagnosed with Parkinson’s in 2009.

The researchers sifted through each person’s claims history to draw up a list of all diagnoses received and medical procedures undergone from 2004 to 2009. With that information, they could develop an algorithm factoring in the use and prevalence of insurance-claim codes related to the common health problems associated with Parkinson’s – including tremors, posture abnormalities, cognitive dysfunction, gastrointestinal issues, and others.

The algorithm also accounted for factors that indicated patients were unlikely to develop Parkinson’s – finding that obesity-related conditions and histories of tobacco smoking, cancer, cardiovascular disease, and certain other ailments were associated with lower likelihood.

A total of 536 diagnosis and procedure codes were factored into the algorithm, along with other risk factors like age, sex, and race/ethnicity. When applied to the data set, the algorithm correctly identified 73% of the people who would be diagnosed with the disease in 2009, and 83% of the people who would not – making it far more effective at predicting Parkinson’s than more basic algorithmic models.

According to a key researcher in the study, the effectiveness rates suggest that the predictive model could be used in AI applications to identify those who will develop the disease years (or even decades) before onset:

“Using this algorithm, electronic medical records could be scanned and physicians could be alerted to the potential that their patients may need to be evaluated for Parkinson’s disease,” said Brad A. Racette, MD, the Robert Allan Finke Professor of Neurology and the study’s senior author.

Racette’s vision for the algorithm would likely require far more research and development, but could potentially make a very strong (and positive) impact across many areas of the healthcare space.

As Racette and his Washington University team continue refining their algorithm, startups interested in using AI coupled with medical data to improve healthcare costs and outcomes will likely be paying attention. Clover Health, for example, is focused on applying data science to preventive care, beginning with analysis of Medicare Advantage claims for the patients Clover insures.

When you consider that approximately 60,000 Americans are diagnosed with Parkinson’s each year (according to the Parkinson’s Disease Foundation), faster and earlier diagnoses of the disease could be beneficial in many different ways.

For example, the researchers found the Parkinson’s patients in their data set typically endured a flurry of doctor visits and medical tests in the 18 months before diagnosis in 2009. If diagnosed earlier, it can potentially lower care cost and improve health outcomes – which can ultimately be beneficial for both patients and insurers.

The original study “A predictive model to identify Parkinson disease from administrative claims data” was published in the journal Neurology in September 2017. Full information is available here.