Abstract

1 Zabludowicz Center for Autoimmune Diseases, Sheba Medical Center, affiliated to Tel-Aviv University, Israel
2 Predicta Med, LTD
3 Maccabi Institute for Research & Innovation, Maccabi Healthcare Services, Tel Aviv, Israel.
4 Sackler Faculty of Medicine, Tel Aviv University, Tel Aviv, Israel.
5 Pediatric Gastroenterology Unit, Dana-Dwek Children’s Hospital, Tel-Aviv Sourasky Medical Center, Tel Aviv,
Israel.

Rheumatoid Arthritis (RA) is an autoimmune disease characterized by inflammatory arthritis. Timely
diagnosis remains a challenge due to non-pathognomonic and overlapping symptoms. Up to 40% of
patients may be initially misdiagnosed

1,2 causing delayed diagnosis and treatment, potentially leading
to increased risk of complications and irreversible joint damage. Machine learning decision support
tools can alert physicians to patients who would otherwise be misdiagnosed, potentially improving
patient outcomes, and reducing costs. A proprietary machine learning algorithm, PredictAITM, was
developed with the aim of identifying RA earlier in the primary care setting.

This retrospective study included ~2.5M electronic medical records (EMR) from Israel’s 2nd largest
HMO, Maccabi Healthcare Services. Sufficient structured data was available between the years 2005
– 2021. Rheumatoid arthritis diagnostic criteria were: age 18 or older, and at least three RA diagnosis
codes (ICD-9: 714.0 or ICD-10:M06.9) by any provider (from separate visits), with at least 1 RA
diagnostic code from a Rheumatologist, within a 2-year period 3. Patients who were subsequently diagnosed with similar rheumatologic conditions were excluded from the case cohort 3. The first date
of an RA diagnosis by any provider was used as the reference event. The algorithm used up to 3 years
of data beginning one year prior to the reference event (e.g., a patient’s first RA diagnosis was in Jan
2015, the model analyzed data from Jan 2011 throughl Jan 2014) to make its prediction.

Of 2,471,267 patients, 340 had a diagnosis of RA (72% female) and available antedating data. With
specificity set to 90%, PredictAITM identified 181 (53%) of these patients 1 year prior to the provider’s
first recorded diagnosis. Discriminatory accuracy area under the curve (AUC) was 84% (Figure 1).

PredictAITM was able to accurately identify an RA diagnosis in 53% of patients before being
diagnosed by a provider. It has the potential to reduce time to diagnosis and therefore expedite
treatment initiation.

Figure 1- Identification of Rheumatoid Arthritis Patients 1 Year Prior to PCP Suspicion, ROC Plot (TPR
vs FPR)

1- Chauhan K, Jandu JS, Goyal A, et al. Rheumatoid Arthritis. [Updated 2021 Oct 7]. In:
StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2021 Jan. Available from:
https://www.ncbi.nlm.nih.gov/books/NBK441999/
2- Gomez D, Saavedra-Martinez G, Villarreal L, et alSAT0108 Misdiagnosis of Rheumatoid
Arthritis – The PhotographyAnnals of the Rheumatic Diseases 2015;74:689.
http://dx.doi.org/10.1136/annrheumdis-2015-eular.2532
3- Widdifield J, Bernatsky S, Paterson JM, Tu K, Ng R, Thorne JC, Pope JE, Bombardier C.
Accuracy of Canadian health administrative databases in identifying patients with
rheumatoid arthritis: a validation study using the medical records of rheumatologists.
Arthritis Care Res (Hoboken). 2013 Oct;65(10):1582-91. doi: 10.1002/acr.22031.