May 2 2024Reviewed by Lexie Corner
In a recent paper published in the journal Patterns, a team of researchers from the Ohio State University created a novel artificial intelligence model that mimics the results of randomized clinical trials to identify the courses of action that heart disease patients should take to avoid stroke.
Ping Zhang. Image Credit: Ohio State University
A foundation model approach, akin to generative AI tools like ChatGPT, was used to front-load the model with de-identified data on millions of patients obtained from healthcare claims information provided by employers, health plans, and hospitals.
Researchers could fine-tune the model with information about particular health conditions and treatments (in this case, stroke risk) by pre-training it on a sizable cache of general data. This allowed them to estimate the causal effect of each therapy and identify the most effective one based on the unique characteristics of each patient.
The team stated that their model performed better than seven other models and produced treatment recommendations that matched those of four randomized clinical trials.
No existing algorithm can do this work. Quantitatively, our method increased performance by 7 % to 8 % over other methods. And the comparison showed other methods could infer similar results, but they can’t produce a result exactly like a randomized clinical trial. Our method can.
Ping Zhang, Study Senior Author and Associate Professor, Department of Computer Science and Engineering and Biomedical Informatics, Ohio State University
Researchers do not intend to replace gold-standard clinical research; rather, they believe machine learning could expedite clinical trials and enable more individualized patient care, thus saving time and money.
Our model could be an acceleratory module that could help first identify a small group of candidate drugs that are effective to treat a disease, allowing clinicians to conduct randomized clinical trials on a limited scale with just a few drugs.
Ruoqi Liu, Study First Author and Ph.D. Student, Department of Computer Science and Engineering, Ohio State University
The team dubbed the suggested framework CURE: Causal Treatment Effect estimation.
Liu highlighted the versatility of a treatment effect estimation model pre-trained with vast amounts of unlabeled real-world data, applicable across numerous diseases and drugs.
“We can pre-train the model on large-scale datasets without limiting it to any treatments. Then we fine-tune the pre-trained model on task-specific small-scale datasets so that the model can adapt quickly to different downstream tasks,” said Liu.
MarketScan Commercial Claims and Encounters from 2012 to 2017 provided the unlabeled data used to pre-train the model. This data included three million patient cases, 9,435 medical codes (including 282 diagnosis codes), and 9,153 medication codes.
Two of Liu's model-building strategies increased CURE's power: pre-training a deep synergized patient data-knowledge foundation model using medical claims and knowledge graphs at scale and filling in gaps in patient records by pairing patient information with biomedical knowledge graphs that represent biomedical concepts and relationships.
“We also proposed KG-TREAT, a knowledge-enhanced foundation model, to synergize the patient data with the knowledge graphs to have the model better understand the patient data,” said Liu, who was the first author of a March Proceedings of the AAAI Conference on Artificial Intelligence paper describing the knowledge graph work.
After further fine-tuning, the model predicts which patient outcomes would correspond to different treatments based on pre-trained data overlapped with more detailed information on medical conditions and therapies. This process yields treatment effect estimates.
Study comparisons with other machine learning tools and validation against clinical trial data revealed that the broad pre-training is the foundation of CURE's efficacy, and the addition of knowledge graphs enhanced the model's performance even further.
Zhang sees a time when doctors will be able to use this kind of algorithm, loaded with data from tens of millions of electronic health records, to access a patient's “digital twin” and use the model as a treatment guide. Of course, this presumes the FDA approves AI as a decision-support tool.
This model is better than a crystal ball: Based on big data and foundation model AI, we can have reasonable confidence to be able to say what treatment strategy is better. We want to put physicians in the driver’s seat to see whether this is something that can be helpful for them when they’re making critical decisions.
Zhang, Study Lead Author and Core Faculty Member, Translational Data Analytics Institute, Ohio State University
Zhang leads the Artificial Intelligence in Medicine Lab.
The National Institutes of Health funded this research. IBM Research’s Pin-Yu Chen was a co-author of the CURE in Patterns study, and Anytime AI’s Lingfei Wu co-authored KG-TREAT in AAAI.
Journal Reference:
Liu, R., et al. (2024) KG-TREAT: Pre-training for Treatment Effect Estimation by Synergizing Patient Data with Knowledge Graphs. Patterns. doi.org/10.1609/aaai.v38i8.28727
Source: https://www.osu.edu/