From Fitbit to first diagnosis: AI is rewriting the RA playbook Save

Artificial intelligence isn’t a distant frontier anymore. It’s here, and it appears able to detect signs of rheumatoid arthritis (RA), possibly before we can clinically detect it. Two studies presented at this year’s ACR meeting highlight just how close we may be to a future where algorithms flag disease before we can and monitor activity with minimal patient burden.

Predicting RA before diagnosis

Ayanian and colleagues from Mayo Clinic (Abstract 2260) used a generative AI model trained on 7.8 million clinical notes to see if it could identify patients with RA months before their clinical diagnosis. Using the Jina-embeddings V3 model, a language model fine-tuned to interpret unstructured medical text, the team analysed 4100 confirmed RA cases alongside nearly 80,000 controls from two decades of Mayo data.

The model read everything: clinic notes, nursing documentation and provider communications. After just 12 hours of training on high-performance GPUs, the system could distinguish future RA patients from non-RA controls with an impressive average precision of 0.8 up to a year before diagnosis.

In practical terms, that means the AI was often able to spot a patient on the RA trajectory months in advance, based purely on routine documentation. As the group at Mayo Clinic note, this could eventually allow primary care physicians or non-rheumatology specialists to receive early alerts suggesting a rheumatology referral, before irreversible joint damage begins.

It’s not hard to imagine the potential implications: automated chart readers that spot evolving RA patterns, integrating antibody results, joint symptoms, or subtle text patterns we might otherwise miss. Of course, as the authors emphasise, this work is preliminary. The model still needs external validation and benchmarking against curated phenotypic data. But it is nonetheless an extraordinary proof-of-concept.

AI at the point of care

While the Mayo group looked at how AI can predict RA onset, Jeffrey Curtis and colleagues (Abstract 1661) explored how AI can optimise disease monitoring once treatment begins.

Their prospective study enrolled 150 RA patients starting adalimumab or upadacitinib and combined patient-reported outcomes (PROs) with passively collected physiologic data from Fitbit devices. Over three to four months, machine-learning models were trained to classify whether patients achieved low disease activity (LDA).

The real surprise was that models using as few as one to three PROs, collected as infrequently as every two weeks, achieved over 80% accuracy in predicting low disease activity. The most informative variables were weekly PROMIS fatigue, pain interference, and RADAI-5 scores. Adding Fitbit data, specifically steps and sleep, only slightly improved accuracy.

In a world increasingly overwhelmed by data collection fatigue, this suggests that a limited set of PROs could reliably reflect clinical state, with or without a wearable device. The AI essentially learned how to translate patient-reported experience into an accurate snapshot of disease activity between visits.

A new model of care?

Viewed together, these studies demonstrate two things:

• AI detecting early RA from the digital footprint of pre-diagnostic clinical notes.
• AI monitoring disease control through minimal, low-burden patient data once therapy starts.

Neither replaces the clinician, but they can assist us.

There are caveats. Both projects used sophisticated but still experimental methods. Privacy, interpretability and integration into EHR systems remain major hurdles. Additionally, there is the key question of, can algorithms trained on curated or institution-specific datasets generalise to diverse clinical populations?

Still, this feels an inevitable (and exciting) direction of travel. Generative AI is no longer a novelty. It’s beginning to read our notes, monitor devices and learn the language of rheumatology. Whether predicting onset or confirming remission, it is perhaps coming a step closer to real-world clinical use.