Using Artificial Intelligence to Find Undiagnosed Generalized Pustular Psoriasis Patients

Introduction & Background

The AAD, with critical support from Boehringer Ingelheim and powered by OM1 launched an innovative breakthrough project to develop education resources to improve the evaluation, diagnosis and treatment of patients suffering from generalized pustular psoriasis (GPP), a rare form of psoriasis. This initiative connects DataDerm’s 54 million patient encounters with OM1’s Patient Finder™ tool to improve understanding of GPP Patients’ journey.

Data Sources

In this study, we characterized two GPP patient cohorts. One (Cohort 1) was drawn from patients diagnosed in AAD’s DataDerm™, a national clinical registry containing real-world dermatology data on more than 54 million patient encounters, covering more than 14 million unique patients. 

The other (Cohort 2) was isolated within the OM1 Real-World Data Cloud™ (RWDC), a large, multisource dataset including combined, normalized, and deterministically linked data drawn from medical and pharmacy claims, electronic medical records, social determinants of disease, death data, and other information. The RWDC includes data on approximately 340 million individuals in the U.S., from 2013 through the present.

OM1 Patient Finder: How it Works

OM1 Patient Finder is an artificial intelligence (AI) application built on PhenOM®, OM1’s digital phenotyping platform. PhenOM isolates patterns in longitudinal patient histories and combines them to create data ‘fingerprints’ – digital phenotypes – associated with characteristics or outcomes of interest. Patient Finder uses this technology to isolate patients with undiagnosed or misdiagnosed disease, based on similarities between their health histories and the target condition’s digital phenotype.

To identify GPP patients, we used patients already diagnosed to isolate a GPP ‘fingerprint’ and calibrate Patient Finder.

We then used Patient Finder to score and rank a test dataset, with patients most likely to have GPP ranked highest. Those in the highest-risk category studied so far – the top 0.1% of the population analyzed – were flagged.

Outcomes

We successfully identified GPP patient cohorts within each source dataset, with 1,883 patients in Cohort 1 and 7,558 in Cohort 2.

Patient Finder performed very well in identifying GPP patients using the source datasets for calibration. In testing, Patient Finder achieved an area under the receiver operating characteristic curve (AUROC) of 0.80. Relative to baseline or ‘background’ GPP prevalence, patients in the top group flagged by Patient Finder - those judged most likely to have GPP – were approximately 40 times more likely to have the condition. In practice, this means that Patient Finder is successfully able to flag patients at highly elevated risk for GPP using widely available real-world data, while these patients are still undiagnosed.

Full Population

Top 0.1%

Gray =non-GPP, orange = GPP; normalized to 1% baseline GPP prevalance

Looking Ahead

With nearly 10,000 GPP patients, this study is providing novel insights into patient journeys that may be helpful in improving these patients’ outcomes in the future. Current Patient Finder performance confirms that real-world data can be used to isolate and apply a GPP-specific digital phenotype, highlighting those at risk. Future work will focus on validating the performance of the application to confirm the same algorithms will perform well across different datasets. Insights drawn from this work will be used to inform provider education around GPP patient journeys and diagnosis. Eventually, Patient Finder can be deployed to proactively flag high-risk patients in clinic. Through these tools, we hope to shorten GPP patients’ time to diagnosis, decrease their suffering, and improve their access to effective treatment.