Abstract Paroxysmal nocturnal hemoglobinuria (PNH) is a rare, chronic, potentially life-threatening hematological disorder that is associated with uncontrolled terminal complement activity leading to intravascular hemolysis, thromboembolic events, organ damage, impaired quality of life, and increased mortality. PNH can be challenging to diagnose due to the diversity and non-specific nature of its clinical manifestations. The diverse and nonspecific clinical symptoms of PNH and its rarity can delay screening and subsequent diagnosis by as much as 10 years. Artificial intelligence (AI) is increasingly being used in diagnostic methods to analyse clinical data and identify patterns that may help improve diagnostic accuracy and identify more that can benefit from treatment. We developed a PNH AI algorithm by integrating knowledge-based expert rules, machine learning, and symptom scoring to identify patients at risk of PNH. This development uses fully anonymized electronic health records (EHRs), with no patient consent required, as approved by the Rzeszow University Ethical Commission (ethical assessment No. 2024/05/025). It assists in the early identification of PNH by analyzing EHRs in a real-world hospital network. In a prospective study across 14 hospitals in Poland, the AI tool screened 1,307,140 patients over a 14-month period and flagged 356 (0.03%) individuals as high-risk for PNH. Physicians reviewed these flagged cases, and 119 high-risk patients were subsequently referred for confirmatory flow cytometry and tested. PNH was diagnosed in 13 of the referred patients, corresponding to a positive predictive value (PPV) of 10.92%. The high-risk cohort identified by the algorithm exhibited distinct clinical characteristics compared to the general screened population. Fatigue and anemia present in 76.4% and 72.2% of flagged patients respectively were the most prevalent symptoms in this high-risk group, each significantly more frequent than in non-flagged patients (both p0.001). Likewise, known PNH-associated conditions were overrepresented among flagged individuals. Nearly half had co-existing myelodysplastic syndrome (MDS, 49.2%), and subsets of patients had a history of aplastic anemia (8.7%) or Budd-Chiari syndrome (1.7%). These conditions were significantly less frequent (0.5%) in the rest of the screened population (p0.001 for each). Notably, the median age at symptom onset for common features like fatigue and anemia was higher in the high-risk group compared to non-flagged patients by 14 and 13 years respectively. The median age of the high-risk PNH cohort was 69.5 years, reflecting the older demographic profile identified by the AI algorithm when compared to median ages ranging from 39 to 45.1 years in previously described cohorts. A diagnostic delay in the cohort of PNH confirmed patients, which might have been preventable with earlier deployment of the AI tool, ranged from 74 to 1337 days. In conclusion, this real-world study demonstrates that an AI-based EHR screening platform can successfully pinpoint previously unrecognized PNH patients with high diagnostic accuracy. To our knowledge, it represents the first large-scale implementation of AI for PNH screening in routine clinical practice. Our findings underscore the potential of integrating machine learning tools into clinical workflows to expedite rare disease diagnosis and enable earlier detection and intervention for conditions like PNH.
Dąbrowski et al. (Mon,) studied this question.