Using Artificial Intelligence to Tackle Cognitive Impairment

Cleveland Clinic, in collaboration with the National Football League Players Association (NFLPA), is launching a research initiative to use artificial intelligence (AI) and machine learning to better characterize neurological disease, with the goal of improving diagnosis, prognosis prediction, and disease intervention and prevention.

The $1.275 million project, funded by the NFLPA, includes the creation of a coordinated research network consisting of multiple bodies drawn from healthcare facilities, research groups and private industry with the aim of developing clinically relevant, user-friendly algorithms based on large datasets.

“Neurological diseases such as Parkinson’s and Alzheimer’s have become public health emergencies as our population ages, and they are also of utmost important to the NFLPA,” says the initiative’s principal investigator, Jay Alberts, PhD, Vice Chair of Innovation for Cleveland Clinic’s Neurological Institute. “This collaborative research project will use machine learning techniques to address fundamental gaps in our understanding of disease processes and management.”

From data to clinical decision tool

The initiative will develop machine learning algorithms related to cognitive impairment, first using general population longitudinal data and then data from a cohort of former NFL players. The process will consist of the following steps:

• Identify various phenotypes of cognitive impairment. These will vary in severity from mild impairment to dementia, and incorporate different forms of dementia.
• Define proxy (matched control) cohorts of patients who are representative of the selected phenotype but who did not develop cognitive impairment.
• Construct trajectory models of the cognitive impairment phenotypes, linking patient-entered data to specific outcomes.
• Model effective treatment interventions — including behavioral, surgical and pharmacological — on disease progression.
• Create clinical decision-support tools to aid early recognition of various diagnoses involving cognitive impairment, as well as predict prognosis and guide interventions.

A rich data source

The first phase will leverage electronic health records from more than 3.5 million patient visits within Cleveland Clinic’s Neurological Institute over the past decade. This database includes thousands of patients treated for Parkinson’s disease, Alzheimer’s disease, multiple sclerosis and amyotrophic lateral sclerosis. Many had been treated at Cleveland Clinic for years before receiving their neurological diagnosis, so their data may contain valuable clues to their disease before a diagnosis was made.

In addition to standard medical record data, many patients’ electronic health record includes routinely completed annual and semi-annual patient-reported outcomes and other patient-entered data reflecting quality of life, depression, cognition and physical function. This information can enable a holistic view of the patient by incorporating medical, cognitive, behavioral, functional and social components of health.

“The amount and consistency of long-term patient outcomes data at Cleveland Clinic are unique in the world,” says Dr. Alberts. “This wealth of data has enabled us to previously create successful models to stratify risk of hospital readmission, high use of outpatient services, vulnerability to iatrogenic harm and overall mortality.”

He adds that one goal of the coordinated research network is to acquire outside data from public, private and clinical trial datasets to further enrich the model.

Key algorithmic attributes: Clarity and usability

According to Dr. Alberts, the models are being built for clinical integration to support physician decision-making. Statisticians, mathematicians and clinical experts will collaborate on developing state-of-the-art, clinically relevant models. The project will be guided by the principles of eXplainable Artificial Intelligence (XAI), which develops AI algorithms with an interface that is more transparent to the user than many existing AI tools.

“The NFLPA is interested in transforming traditional ‘black box’ machine learning models into ‘glass box’ models, without sacrificing performance,” Dr. Alberts explains. “The goal is for clinicians and other users to understand why the algorithm comes to a conclusion, increasing their ability to use the tool and enhancing their trust in the process.” Moreover, use of this glass box approach helps ensure that models are free from racial or sex bias.

Multiple clinical applications

An ultimate goal of the initiative is for the created algorithms to support earlier diagnosis of neurological syndromes, enable better prediction of disease course and help guide interventions. They could also help identify better candidates for drug trials. Dr. Alberts expects the models to be refined as more treatments become available and their outcomes are documented.

He anticipates that the project will lead to clinically relevant algorithms within a year. They will be published and be of immediate use to the NFLPA and Cleveland Clinic.

“The resulting clinical decision-support tools should be especially useful to busy community primary care providers who can employ them to better care for neurological patients or trigger appropriate referral to specialists in neurological disease,” Dr. Alberts concludes. “The tools will also help inform education and safety initiatives for the NFLPA.”