Implementation of the genome-informed risk assessment (GIRA) may lead to large disruptions to the health system

This study evaluating the Genome-Informed Risk Assessment (GIRA) in the Penn Medicine Biobank reveals that while the tool effectively identifies a large proportion of high-risk patients across diverse ancestries, its implementation at scale poses significant challenges due to disparities in risk stratification by ancestry and socioeconomic status, as well as a notable gap between its ability to predict prevalent versus incident disease cases.

The Gist

Imagine your body is a complex house. For years, doctors have tried to predict if the roof might leak (heart disease), if the pipes might burst (diabetes), or if the foundation might crack (cancer).

Traditionally, they've looked at two things:

1. The Family Album: Did your parents or grandparents have these problems? (Family History)
2. The Current Inspection: Do you smoke? Are you overweight? Do you have high blood pressure? (Clinical Factors)

Recently, scientists invented a new tool called GIRA (Genome-Informed Risk Assessment). Think of GIRA as a super-powerful blueprint scanner. It doesn't just look at the current state of the house; it reads the original architectural plans (your DNA) to see if there are hidden structural weaknesses you were born with.

This paper is like a real-world stress test for that blueprint scanner. The researchers took this new tool, which was originally tested in a specific research lab (eMERGE), and tried it out in a massive, real-world hospital system in Philadelphia (Penn Medicine) to see if it actually works for regular people.

Here is what they found, broken down into simple concepts:

1. The "High-Risk" Crowd is Huge

When they scanned the blueprints of nearly 48,000 people, they found something surprising: Half of them (50%) were flagged as "High Risk" for at least one of nine major diseases.

• The Analogy: Imagine walking into a stadium and a scanner beeps at half the crowd, saying, "Hey, your house might have a leak!"
• The Problem: If a hospital actually used this tool on everyone, it would be a massive disruption. The doctors would be overwhelmed trying to treat or monitor 50% of their patients. It's like trying to fix every roof in a city at once; the system might collapse under the weight.

2. The Scanner Isn't Perfect for Everyone (The Ancestry Issue)

The scanner was calibrated mostly using blueprints from people of European descent. When they tested it on people of different backgrounds, the results were uneven.

• The Analogy: Imagine a metal detector tuned to find gold. It works great in a gold mine (European ancestry), but if you take it to a copper mine (African ancestry), it might beep too often (false alarms) or miss the copper entirely.
• The Finding: People of African ancestry were flagged as "high risk" much more often than others, not necessarily because they were sicker, but because the scanner wasn't perfectly tuned for their specific genetic "blueprints." Meanwhile, people of Asian ancestry were flagged less often, partly because the scanner relies heavily on family history, which can be hard to track accurately in electronic records.

3. The "Crystal Ball" vs. The "Rearview Mirror"

This is the most critical finding. The scanner is great at looking in the rearview mirror (identifying people who already have a disease) but is a bit shaky as a crystal ball (predicting who will get sick in the future).

• The Analogy:
  • Prevalent (Rearview Mirror): The scanner correctly identified that if you have a leaky roof, you probably have water damage right now. It was very accurate here.
  • Incident (Crystal Ball): When they tried to predict who would get a leak next year, the scanner was much less confident. It said, "You might get a leak," but the actual number of people who got sick was lower than predicted.
• The Takeaway: Just because the blueprint says you are at risk doesn't mean the roof will leak tomorrow. Other factors (like the weather, how well you maintain the house, or your socioeconomic status) play a huge role.

4. The "Weather" Matters (Social Factors)

The study found that people living in more deprived areas (lower income, less access to care) were flagged as "high risk" more often.

• The Analogy: The scanner is picking up that the house is in a stormy neighborhood. It's not just the blueprint; it's the environment. If you don't have money to fix a small crack, it becomes a big problem. The scanner is detecting the risk created by poverty and lack of access, not just the DNA.

The Bottom Line

The GIRA tool is a promising new gadget, but it's not ready to be handed out to every patient in every hospital tomorrow.

• It works well for identifying people who already have hidden risks.
• It needs tuning to work fairly for people of all racial backgrounds.
• It needs context. You can't just look at the DNA blueprint; you have to look at the neighborhood, the weather, and the family history to know if the roof will actually leak.

If we implement this too quickly without fixing these issues, we risk overwhelming the healthcare system with false alarms or missing the people who actually need help. It's a powerful tool, but like any new technology, it needs to be calibrated before it can save the day.

Technical Summary

1. Problem Statement

The Genome-Informed Risk Assessment (GIRA), developed by the eMERGE Network, represents a standard approach for implementing genomic precision medicine at scale by combining polygenic risk scores (PGS), monogenic risk, and family history. While eMERGE validated GIRA within its own multi-site cohorts, there is a critical gap in understanding how GIRA performs in independent healthcare systems with different patient demographics, ancestry compositions, and environmental factors.

Key challenges addressed include:

• Generalizability: Does GIRA maintain predictive accuracy in populations outside the eMERGE network (specifically the University of Pennsylvania Health System)?
• Prevalent vs. Incident Prediction: Most prior validations focused on prevalent cases (existing disease). There is a lack of longitudinal data on whether GIRA effectively predicts incident cases (future disease onset), which is crucial for preventative interventions.
• Equity and Disparities: How does GIRA perform across diverse ancestry groups (e.g., African/African American vs. European) and socioeconomic strata, given known biases in polygenic scores?
• Systemic Impact: What is the burden on the healthcare system if a large proportion of the population is flagged as "high-risk"?

2. Methodology

The study utilized the Penn Medicine Biobank (PMBB), an electronic health record (EHR)-linked biobank embedded within the University of Pennsylvania Health System (UPHS).

• Cohort: 48,279 adult participants (aged 18–75) with genomic and exome data. The cohort is notably diverse, with 23.2% self-identifying as African/African American (AFR), the largest representation of AFR individuals in any single-institution US medical biobank.
• Conditions: Nine adult-onset conditions were analyzed: Atrial Fibrillation (AFIB), Breast Cancer (BC), Chronic Kidney Disease (CKD), Coronary Heart Disease (CHD), Hypercholesterolemia (HC), Obesity, Prostate Cancer (PC), Type 2 Diabetes (T2D), and Colorectal Cancer (CC).
• GIRA Implementation:
  • Polygenic Risk Scores (PGS): Used the same ancestry-calibrated PGS models and thresholds as eMERGE.
  • Monogenic Risk: Screened for rare, pathogenic variants in 16 genes associated with the conditions.
  • Family History: Extracted from EHR data for specific conditions.
  • Risk Stratification: Participants were labeled "High Risk" if they met criteria for any of the three components (PGS, Monogenic, or Family History).
• Phenotyping: Validated algorithms from PheKB, CCW, and PRIMED were used to define cases.
  • Prevalent Cases: Diagnosed before, at, or after biobank entry.
  • Incident Cases: Diagnosed strictly after biobank entry (longitudinal follow-up).
• Statistical Analysis:
  • Prevalence: Logistic regression to calculate Odds Ratios (OR).
  • Incidence: Cox proportional hazard models to calculate Hazard Ratios (HR) and 15-year cumulative incidence.
  • Stratification: Analyses were stratified by ancestry (EUR, AFR, etc.) and Social Deprivation Index (SDI) quartiles.

3. Key Contributions

• First Independent Validation: This is the first study to validate GIRA in a healthcare system independent of eMERGE, specifically focusing on a diverse, real-world patient population.
• Longitudinal Assessment: It provides the first evaluation of GIRA's ability to predict incident disease, contrasting it with prevalent disease prediction.
• Ancestry-Specific Performance: It offers granular data on how GIRA performs in AFR populations compared to EUR populations, highlighting specific conditions where calibration is needed.
• Systemic Impact Quantification: It quantifies the potential volume of high-risk patients a health system would need to manage, highlighting the operational challenges of scaling genomic medicine.

4. Key Results

A. High-Risk Prevalence and Demographics

• High-Risk Rate: 50.1% (n=24,185) of the PMBB cohort was classified as high-risk for at least one condition.
  • 30.4% were high-risk due to polygenic and/or monogenic factors.
  • 28.8% were high-risk due to family history.
• Ancestry Disparities: Significant differences were observed in high-risk proportions:
  • AFR: 56.6% (significantly higher than the overall average).
  • East Asian (EAS): 42.0%.
  • South Asian (SAS): 40.0%.
  • Note: The lower rates in Asian populations were driven primarily by family history reporting patterns rather than genetic risk estimates.
• Socioeconomic Factors: The lowest quartile of the Social Deprivation Index (most deprived) had a higher high-risk rate (53.7%), driven by PGS and family history, suggesting environmental factors influence risk classification.

B. Predictive Accuracy: Prevalent vs. Incident

• Prevalent Cases: GIRA performed well, identifying prevalent cases with Odds Ratios (OR) ranging from 2.28 to 3.97, consistent with eMERGE findings.
• Incident Cases (Longitudinal): Predictive power was substantially attenuated for incident cases.
  • For 5 of the 9 conditions, the Hazard Ratios (HR) were significantly lower than the ORs.
  • Example (Atrial Fibrillation): OR was 2.36, but HR dropped to 1.31.
  • Example (CKD): OR was 3.63, but HR dropped to 1.82.
  • This suggests that while GIRA identifies those who have or had the disease well, it is less effective at predicting who will develop the disease in the future compared to prevalent associations.

C. Ancestry and Condition-Specific Variability

• AFR Performance: GIRA showed reduced performance in AFR individuals for several conditions (AFIB, CHD, T2D, Obesity), where confidence intervals often did not reach the OR > 2 threshold.
• Demographic Mismatch: The demographic composition of "High Risk" individuals differed from the actual incident cases for some conditions.
  • Example (AFIB): High-risk individuals were enriched for European ancestry, whereas incident AFIB cases were enriched for AFR ancestry. This indicates a potential miscalibration of genetic risk scores for AFR populations regarding future disease onset.

D. Systemic Implications

• The finding that ~50% of the population is flagged as high-risk suggests that implementing GIRA at scale could lead to large disruptions in the health system due to the sheer volume of patients requiring follow-up, counseling, and preventative interventions.

5. Significance and Conclusion

The study concludes that while GIRA is a robust biomarker for stratifying prevalent disease risk and is largely generalizable across independent healthcare systems, its clinical utility for incident disease prediction is more limited than previously suggested by prevalent-only studies.

• Clinical Translation: The attenuation of risk estimates in longitudinal data (HRs) indicates that clinicians must be cautious not to overestimate future risk based solely on prevalent associations.
• Equity: The performance gaps in AFR populations and the influence of social deprivation highlight the need for condition-specific and ancestry-specific calibration before widespread implementation to ensure equitable care.
• Operational Reality: The high proportion of the population flagged as "high-risk" (50.1%) poses a significant logistical challenge. Health systems must develop scalable workflows to manage this volume without overwhelming resources or causing "alert fatigue."

Ultimately, the paper advocates for careful, context-specific validation and refinement of genomic risk tools before they are deployed broadly in diverse, real-world healthcare settings.