Deep learning algorithm shows promise for predicting RA outcomes
medwireNews: Findings from a proof-of-concept study suggest that deep learning models based on electronic health records may be a feasible approach for forecasting disease outcomes in people with rheumatoid arthritis (RA).
“Given the volume of data available in electronic health records (EHRs), the number of possible treatment and outcome trajectories associated with heterogeneous patient comorbidities, medications, and other factors is greater than a human, even an experienced physician, can use,” explain Atul Butte, from the University of California in San Francisco, USA, and fellow researchers.
They add that “[d]eep learning, a subdiscipline of artificial intelligence, has [...] demonstrated multiple successes in clinical applications,” including diabetic retinopathy screening, classification of cardiovascular arrhythmias, and predicting mortality risk in hospitalized patients.
As reported in JAMA Network Open, Butte and colleagues developed three models using data from the electronic health records of 820 RA patients from two different healthcare systems: a university hospital (n=578) and a safety-net hospital (n=242). Variables included in the models were prior CDAI score, erythrocyte sedimentation rate, C-reactive protein levels, the presence of autoantibodies, demographics, and the use of DMARDs and glucocorticoids.
The best-performing model – taking into account 1 year of each patient’s history – was trained in the university hospital cohort and demonstrated “excellent forecasting performance” in that cohort, say the study authors.
Indeed, the model correctly predicted whether patients were in remission, or had low, moderate, or high disease activity according to their CDAI score on 91% of occasions.
Butte and team say that this university hospital cohort-trained model also “performed well” when applied to the safety-net cohort, correctly predicting patient outcomes on 74% of occasions. They note that this good predictive ability occurred “despite marked differences in the patient populations” in the two healthcare systems; for example, individuals in the university hospital cohort visited their rheumatologist more often than those in the safety-net cohort (median time between visits 100 vs 180 days) and were more likely to be prescribed biologics (63.0 vs 28.9%).
The researchers concede, however, that the model’s performance in the safety-net cohort “was poorer compared with the first health system.”
Looking to the future, Butte and colleagues say that “models built from large pooled patient populations may be the most accurate, giving everyone access to the most robust models trained on the largest and most diverse patient populations possible.”
And they conclude: “Model performance is nearing the point at which models are good enough to warrant launching a prospective clinical trial to evaluate their usefulness in aiding practitioners and patients to prognosticate RA outcomes or simulate outcome trajectories under different treatment scenarios.”
medwireNews is an independent medical news service provided by Springer Healthcare. © 2019 Springer Healthcare part of the Springer Nature group