AI machine learning algorithm useful in early detection of PsA

A significant lag in diagnosis of psoriatic arthritis (PsA) may result in joint destruction. A machine learning algorithm proved to be a valuable tool to detect PsA up to 4 years prior to initial diagnosis.

PsA is a good candidate to develop a machine learning algorithm, especially for dermatologists, since it is a prevalent disease, particularly among psoriasis patients, where skin psoriasis usually precedes PsA. Moreover, early diagnosis and treatment can prevent irreversible joint damage and disability. The dermatologist is often the first physician to suspect PsA. “This is why such an algorithm is important for us,” Dr Jonathan Shapiro (Maccabi Healthcare Services, Israel) emphasised [1]. Artificial intelligence (AI)-based decision support tools are urgently needed in this indication, as 10 to 15.5% of psoriasis patients suffer from undiagnosed PsA, and median time to diagnosis is still more than 2 years [2,3]. In addition, rheumatologists are often unavailable.

In a retrospective study, the performance of the machine learning tool PredictAI was evaluated regarding its ability to identify undiagnosed PsA patients 1‒4 years prior to the first suspicion of PsA (reference event) [1]. Data was analysed from the 2.5 million members of the Maccabi Healthcare Services Study population. All diagnoses of PsA and psoriasis in the adult population between 2008‒2020 were evaluated in the study. As Dr Shapiro explained, the researchers did what is done with all machine learning algorithms: “You take the data you have and divide it into 2 sets: a large set for training the algorithm, and a small set used to train and test your algorithm.” Since prevalence of PsA is higher among psoriasis patients compared with the general population, a general population cohort and a psoriasis cohort were formed. The reference event was the first registered diagnosis of PsA by any physician. The algorithm was trained on 4 years of consecutive data preceding the diagnosis.

In the general population cohort, a cut-off point of 99% was chosen compared with a cut-off point of 90% in the psoriasis cohort. In the general population cohort, PredictAI achieved a sensitivity range of 32‒42% and a population positive predictive value (PPV) range of 10‒8%. In the psoriasis cohort, undiagnosed PsA patients were identified by the algorithm 1 to 4 years before the reference event with a sensitivity range of 38‒50% and a PPV range of 30‒34%.

Parameters the algorithm found most prominent to help make the prediction whether a patient has or has not undiagnosed PsA were number of psoriasis diagnoses, number of arthralgia diagnoses, and intra-articular injections. Parameters which were not associated with undiagnosed PsA were age and number of eczema diagnoses.

“The closer you move to the time of the reference event (the diagnosis of PsA) the more the performance of the algorithm improves,” Dr Shapiro explained. In the psoriasis cohort, PredictAI identified undiagnosed PsA in up to 51% of patients up to 4 years prior to primary care physician´s initial suspicion, potentially reducing time to diagnosis and treatment. Therefore, AI models may play an important future role to improve patient outcomes in this disease.

1. Shapiro J. A machine learning tool for the early identification of undiagnosed psoriatic arthritis patients: is it possible? FC06.02, EADV Congress 2022, Milan, Italy, 7‒10 September.
2. Villani AP, et al. J Am Acad Dermatol. 2015;73:242‒8.
3. Karmacharya P, et al. J Rheumatol. 2021:48:1410‒6.

Copyright ©2022 Medicom Medical Publishers



Posted on 5 November 2022