An artificial intelligence (AI)-based screening approach was able to identify suspected Pompe disease (PD) patients from electronic health records (EHRs), according to a study published in Frontiers in Neurology.

The proprietary AI developed by Symptoma flagged 104 patients out of 350,116 EHRs that it was able to scan in a 1-month time period. Generalist physicians (GPs) reviewed these flagged records and determined that 5 should be categorized as “diagnosed,” 10 as “suspected,” and 7 as “reduced suspicion.” Of these 22 patients, specialist physicians (SPs) determined that 5 were “definite,” 2 were “probable,” 6 were “possible,” 6 were “inconclusive,” and 3 were “unlikely,” resulting in 19 who were clinically plausible for PD.

Read more about PD diagnosis

Continue Reading

Considering the patients removed from suspicion by the GPs and SPs, the AI-based approach had a specificity of 21.15% and 18.27%, respectively. Using this information, the prevalence of PD was estimated to be 1 in 18,427.16 in the EHR dataset provided by the University Hospital Salzburg clinic group comprising 5 hospitals.

Roughly 5.47 patients identified by the AI needed to be screened manually to identify 1 suspected patient, drastically improving screening efficiency. In the EHR dataset, 5 patients had confirmed PD and were detected by the AI and correctly classified by both the GPs and SPs. This information suggests a very low false negative rate using this approach.

“This study shows how an AI-based approach analyzing retrospective EHRs results in resource-efficient identification and automated phenotyping of Pompe disease patients,” the authors said.

“This methodology can potentially improve both the timing and accuracy of identifying rare disease patients,” the authors continued.

Further analysis of patient characteristics found that scapula alata, muscle weakness, myalgia, muscle hypotonia, and myopathy were the characteristics most useful for differentiating EHRs flagged as negative, which means either rejected by GPs or labeled as “unlikely” by SPs (n=85), or from those flagged as positive, which means all remaining (n=19).

Using principal component analysis, found that the third principal component was able to drive a weak separation between positive and other groups. This principal component had the largest contributions from myalgia, dyspnea, and fatigue.


Lin S, Nateqi J, Weingartner-Ortner R, et al. An artificial intelligence-based approach for identifying rare disease patients using retrospective electronic health records applied for Pompe disease. Front Neurol. Published online April 21, 2023. doi:10.3389/fneur.2023.1108222