Using AI to Predict Disease Progression

WHILE ARTIFICIAL intelligence (AI) has been used to support many aspects of healthcare for a number of years, its reliability in clinical decision support (CDS) has grown in the past five to 10. AI can democratize healthcare, providing rural and urban, low-income and affluent areas with the potential to access the same information regardless of location.

AI models train on peer-reviewed literature and high-quality research to help with decision-making support. They transcribe conversations by listening during patient appointments; assist with scheduling, billing and prescription reminders; and now, more than ever, support disease diagnosis and targeted treatment plans, learning and perfecting its knowledge from a multitude of data points. In this new era of medicine, exciting opportunities abound as AI continues to integrate into practice settings.

The possibilities of AI are particularly dramatic in the area of autoimmune disease research where accuracy of disease diagnosis and now disease progression are coming into the fold. From rheumatoid arthritis (RA) and lupus to Alzheimer’s disease and diabetes, AI’s analysis and decision-making support for conditions with known and suspected malfunctions of the immune system are indisputable.

But, despite the incredible potential of AI, some warn of inconsistencies and sometimes inaccurate results produced by machine learning. The ability to adapt to multimodal learning is not always accurate. Reliability, flexibility and contextualization in the face of incomplete and ambiguous information stunts machine learning, making for a stark reminder that overreliance on AI without human intervention could be risky if not downright dangerous.

Even so, doctors and patients continue to demand faster and more accurate diagnoses and disease treatments using AI as an incredibly powerful tool. Continually learning and advancing, this CDS tool will forever stay in the physician’s toolbox as it advances healthcare beyond diagnoses and treatments and into disease progression prediction.

Improving Autoimmune Disease Progression Prediction

Researchers across the U.S. are conducting AI studies to improve disease progression for autoimmune diseases. Results show AI’s abilities in the broader ecosystem of health research and implementation of more effective therapies.

A team of researchers at Penn State College of Medicine conducted a study to pioneer the use of AI in determining who is most at risk for advanced autoimmune disease progression. Through analysis of electronic health records (EHRs) and large genetic studies, a newly created genetic prediction score (GPS) was found to be up to 1,000 percent more accurate than 20 other commonly used methods of disease prediction scoring.

In this study, the researchers used transfer learning to develop its GPS, testing both in patients with RA and lupus. They found AI was able to detect antibodies of those with RA five years before symptoms began. This ability to foretell preclinical disease enables the initiation of suitable and personalized treatments, thus extending patients’ health span and potentially slowing disease progression.

With transfer learning, the system initially trained on smaller and simpler data subsets (for example, distinguishing between a cat and a dog) before it moved on and fine-tuned its skills to learn a different but related set of data (such as malignant and benign tumor cells). Given the smaller subset of patients who have or will develop autoimmune disease (eight percent of the American population, according to the National Institutes of Health, most of whom are women), the real benefit of this transfer learning was its ability to create a reliable forecasting model given a smaller sample size.

In this case, the GPS trained on genome-wide association studies and EHR-based biobanks containing information about genetic variants, lab tests and more to identify genetic differences in those with autoimmune diseases. By determining which patients were in the preclinical stage of autoimmune disease, the machine could create a prediction score for those who would be most at risk of progressing to the disease stage.¹

In another study, researchers at Boston University built an AI model that predicts the presence of amyloid beta and tau proteins using memory score tests and brain scans in combination with age, health history and genetic information to predict a likelihood of developing Alzheimer’s disease. This model not only predicted Alzheimer’s, but also identified the area in the brain that would be affected and could detect disease progression from mild cognitive impairment to Alzheimer’s within a six-year window. Better staging and subtyping of Alzheimer’s enables the development of personalized treatment plans that could potentially slow the disease progression. The model also scored well when tested on participants who were not part of the training regimen, accurately predicting high amyloid or tau levels as an indicator of cognitive decline.²

Type 2 diabetes mellitus, with suspected ties to the immune system, is also benefiting from AI research, with some models achieving up to 90 percent accuracy in predicting diabetes onset before clinical symptoms appear. Another notable highlight was a more accurate diagnosis of diabetic retinopathy using retinal images and clinical data. By using new AI-powered diagnostic technologies that read data from metal oxide semiconductor sensors, breath analysis and real-time retinal screening, patients may one day no longer need invasive blood draws or glucose tolerance tests. The modeling also hinted at the possibility of expansion into prediction for diabetic retinopathy, peripheral neuropathy and others. When these technologies are ready for prime-time U.S. Food and Drug Administration (FDA) review, they may well offer a comprehensive approach that integrates EHRs, noninvasive monitoring and a better ability to manage disease outcomes.³

One of the most promising areas of AI use in prediction medicine is the creation of treatment plans and forecasting of individual response to personalized therapies. For example, in the case of type 1 and type 2 diabetes, AI is enabling earlier diagnoses and better optimization of disease management to help to prevent or delay complications and adverse outcomes. Researchers in one study applied logistic regression analysis to two datasets, one for Pima Indians (type 1) and another for rural African Americans (type 2), with accurate disease prediction of 78 percent and 93 percent, respectively. Patients identified to be at risk were encouraged to adopt healthier behaviors supported by data-driven tools used in predictive modeling for lifestyle adjustments.⁴

The Challenges of AI Decision-Making

While AI is well-known for passing medical exams and text modality challenges, multi-modality inputs and even the data itself can sometimes challenge AI decision-making. For example, inconsistencies in the quality of EHRs may significantly limit machine learning accuracy that may cause fragmented or incomplete data quality. Without human oversight to verify results, health disparities and inequitable treatments may perpetuate.

Stress tests reveal some limitations in multimodal AI robustness and performance. While high scores on medical benchmarks may suggest clinical competence, when input prompts are altered or reordered, such as removing or adding an image, the system can become confused and generate incorrect responses, but they do so with a high confidence of accuracy. More worrisome to some is AI’s sometimes flawed reasoning for determining its response, hinting to competency gaps in multimodal systems capabilities.⁵

The adoption and integration of multi-disease diagnostic AI models into clinical workflows faces both operational and integration obstacles. Effective implementation of AI relies on the ability to pivot to new data collection methods, new analyses based on AI inputs and the capacity to interpret AI’s decision-making and reasoning.

Further, it is imperative that AI tools perform accurately and consistently across a variety of patient populations, infrastructure differences and varying robustness of EHR data so AI modeling does not exacerbate healthcare inequities or bias.

Patchwork Regulatory Environment

The regulatory framework around the use of AI in healthcare is patchy at best. For instance, in the U.S., FDA regulates algorithm-based CDS when it meets the definition of a medical device. But, state laws and executive orders, including those focused on privacy, make the situation more complex. Additionally, the World Health Organization (WHO) in 2023 created a set of expectations for using AI in health that focused on transparency, documentation, risk management, validation, data quality, data privacy and protection, and collaboration. The European Union, also in 2023, created the EU AI Act, “a binding worldwide regulation” for the use of AI in the EU. Both WHO and EU frameworks allow for monitoring of CDS solutions throughout a product lifecycle, ensuring achievements in quality and ethics.⁶ 

Ironically, the benefit of AI’s evolutionary learning is also one of its biggest liabilities from a regulatory perspective. As AI trains on ever-evolving historical data, incoming data that facilitates new learning is subject to privacy laws that make it more difficult to acquire and use the data. FDA has been looking into the potential of synthetic data, generated through computer simulations, to supplement real-world datasets to protect data privacy. Federated learning, a decentralized AI approach, offers another option. Instead of each system collecting and storing large datasets, in federated learning, AI would train across datasets that remain at their original source locations and would, in theory, provide better patient privacy.³ 

Also, differing AI applications require unique regulatory frameworks, yet across the board, regulatory agencies are demanding that AI be explainable, transparent, traceable and accountable. 

Ensuring Accurate and Reliable Prediction

AI’s powerful support of healthcare continues to grow and evolve, creating more robust opportunities to understand disease and design meaningful projection models that forge new treatment outcomes. While there are certainly risks with the use of AI, as with any newer technology, thoughtful design and regulatory and human oversight will help to ensure machine learning supports accurate and reliable disease progression prediction now and into the future.