An algorithm to identify patients with rare genetic disorders and its real-world data application

Bryn D. Webb, Lisa Y. Lau, Despina Tsevdos, Ryan A. Shewcraft, D Corrigan, Lisong Shi, Seungwoo Lee, Jonathan Tyler, Shilong Li, Zichen Wang, Gustavo Stolovitzky, Lisa Edelmann, Rong Chen, Eric E. SchadtmedRxiv 2023 · DOI

TL;DR

Researchers have developed PheIndex, a rule-based digital phenotyping algorithm that sifts through Electronic Medical Record (EMR) data to identify infants and toddlers at high risk for rare genetic disorders. Beyond just identification, the study demonstrates that children of mothers who underwent expansive carrier screening (100+ genes) show significantly better health outcomes and lower "risk scores" than those whose mothers had minimal screening.

What Problem Does This Paper Address?

Identifying rare genetic disorders is a notorious "diagnostic odyssey." While there are over 7,000 known Mendelian disorders, they are individually rare, making them easy to miss in a general pediatric setting.

The authors highlight a major technical hurdle: Electronic Medical Records are poorly equipped for rare disease surveillance. Only about 11–21% of these disorders have specific ICD (billing) codes. Furthermore, clinical data is often trapped in unstructured formats like PDFs or free-text notes, and doctors frequently use "rule-out" codes that create noise in the data. This paper aims to move beyond specific disease-seeking algorithms toward a global metric that flags children based on healthcare utilization patterns—essentially identifying the "clinical footprint" of a genetic disorder before a formal diagnosis is even made.

Methods

The team developed PheIndex, a rule-based algorithm utilizing 13 clinical criteria extracted from structured EMR data. These criteria are weighted as "Major" (3 points) or "Minor" (1–2 points):

  • Major Criteria: Prolonged NICU stays (for term babies), multiple hospitalizations, heart surgeries, metabolic disease codes, and in-hospital death.
  • Minor Criteria: Visits to multiple specialists (e.g., genetics, neurology), frequent ER visits, feeding support (G-tubes), respiratory support, imaging (CT/MRI), and developmental delay codes.

The algorithm produces two outputs:

  1. PheIndex Score: A continuous value (0–24) representing the severity of the clinical burden.
  2. PheIndex Classification: A binary "positive/negative" flag based on specific combinations of major and minor criteria.

To validate the tool, a pediatrician and a clinical geneticist performed a blinded chart review of 200 cases to compare the algorithm's "digital phenotype" against real-world clinical judgment.

Datasets

The study utilized a massive de-identified dataset from the Mount Sinai Health System (MSHS):

  • Cohort: 93,154 mother-child pairs delivered between 2007 and 2019.
  • Follow-up: All children had at least one year of follow-up data, with many tracked up to age three.
  • Carrier Screening Data: The researchers linked the children's outcomes to their mothers' carrier screening (CS) history, categorized into Small (CS-S, ≤4 genes), Medium (CS-M), and Large (CS-L, ≥100 genes) panels.

Key Insights and Ideas

  • Utilization as a Proxy for Genetics: The most powerful insight is that we don't need a specific genetic test result to know a child is at risk. High healthcare utilization—seeing five different specialists or having multiple prolonged hospital stays—is a "phenotypic signature" that can trigger an earlier referral to a geneticist.
  • The Power of Expanded Screening: The study provides real-world evidence for the benefit of expanded maternal carrier screening. Children in the CS-L group (large panels) were 36% less likely to be classified as "high risk" by the algorithm.
  • Proactive vs. Reactive Medicine: Interestingly, children in the CS-L group who did meet the risk criteria reached specialist visits and diagnostic testing earlier than those in the CS-S group. This suggests that parental awareness of carrier status leads to faster clinical action once symptoms appear.

Key Results

  • High Accuracy: In the validation cohort, PheIndex achieved 90% sensitivity, 97% specificity, and 94% accuracy compared to expert chart review.
  • Prevalence: The algorithm flagged 1,088 children (1.2% of the total cohort) as being at high risk for a genetic disorder.
  • Clinical Impact: Children of mothers who had large-panel carrier screening had significantly lower average PheIndex scores (0.57) compared to those with small-panel screening (0.70) at one year of follow-up ($p < 0.001$).
  • Reduced Burden: The CS-L group showed a decreased need for multiple ER visits and specialist consultations over time compared to the CS-S group.

Limitations and Future Work

  • Structured Data Only: The current version of PheIndex relies only on structured data (codes and orders). The authors admit that incorporating unstructured clinical notes (via NLP) would likely catch cases where documentation is incomplete.
  • Geographic Bias: The data comes from a large, diverse metropolitan health system. Results might differ in rural areas where access to specialists (a key metric in the algorithm) is limited.
  • The "Why" of CS-L: While the correlation between large screening panels and better outcomes is strong, the study cannot perfectly account for all confounders, such as the "healthy volunteer" effect (parents who opt for more testing may be more proactive in other areas of healthcare).
  • Future Direction: The ultimate goal is to integrate PheIndex into live EMR systems as a clinical decision support tool, alerting pediatricians to order genetic testing years earlier than they currently do.

Summary updated

Abstract

Abstract Objectives Develop a digital phenotyping algorithm (PheIndex) using electronic medical records (EMR) data to identify children aged 0-3 who have been diagnosed with genetic disorders or present with illness with an increased risk for genetic disorders from a mother-child cohort. Methods We established 13 criteria for the algorithm where two metrics – a quantified score and a classification – were derived. The criteria and the classification were validated by chart review from a pediatrician and clinical geneticist. To demonstrate the utility of our algorithm in real-world evidence applications, we examined the association between size of carrier screening panel (small/≤4 genes [CS-S] vs large/≥100genes [CS-L]) undertaken by mothers prior to delivery, and children classified as presenting with illness with an increased risk for genetic disorders by our algorithm. Results The PheIndex algorithm identified 1,088 such children out of 93,154 live births and achieved 90% sensitivity, 97% specificity, and 94% accuracy by chart review. We found that children whose mothers received CS-L were less likely to be classified as presenting with illness with an increased risk for genetic disorders and a decreased need to have multiple specialist visits and multiple ER visits, compared to children whose mothers received CS-S. Conclusions The PheIndex algorithm can help identify when a rare genetic disorder may be present, and has the potential to improve healthcare delivery by alerting providers to consider ordering a diagnostic genetic test and/or referring a patient to a medical geneticist or other specialists. Article Summary Algorithm using EMR data to identify children who have been diagnosed with a genetic disorder or present with illness with increased risk of genetic disorders. What’s known on this subject With over 7000 Mendelian disorders, identifying children with a specific rare genetic disorder diagnosis through structured EMR data is challenging given incompleteness of records, inaccurate medical diagnosis coding, as well as heterogeneity in clinical symptoms and procedures for specific disorders. What this study adds We developed a digital phenotyping algorithm using electronic medical records (EMR) data to identify children aged 0-3 who have been diagnosed with genetic disorders or present with illness with an increased risk for genetic disorders from a mother-child cohort.
BibTeX
@misc{webb2023algorithm,
  author = {Bryn D. Webb and Lisa Y. Lau and Despina Tsevdos and Ryan A. Shewcraft and D Corrigan and Lisong Shi and Seungwoo Lee and Jonathan Tyler and Shilong Li and Zichen Wang and Gustavo Stolovitzky and Lisa Edelmann and Rong Chen and Eric E. Schadt and Li Li},
  title = {{An algorithm to identify patients with rare genetic disorders and its real-world data application}},
  year = {2023},
  doi = {10.1101/2023.01.27.23285056},
  howpublished = {medRxiv},
}
Figure