Machine Learning to Predict Osteoporotic Fracture Risk from Genotypes

Vincenzo Forgetta, Julyan Keller‐Baruch, Marie Forest, Audrey Durand, Sahir Bhatnagar, John P. Kemp, John A. Morris, John А. Kanis, Douglas P. Kiel, Eugène McCloskey, Fernando Rivadeneira, Helena Johannson, Nicholas C. Harvey, Cyrus Cooper, David M. Evans, Joëlle Pineau, William D. Leslie, Celia M.t. Greenwood, J. Brent RichardsbioRxiv 2018 · DOI

TL;DR

This paper introduces a machine learning model that uses genetic data to predict osteoporotic fracture risk, achieving performance comparable to traditional methods like bone density scans. This "genomic prediction" could significantly improve the efficiency of osteoporosis screening by identifying individuals who truly need further testing, potentially reducing healthcare costs.

What Problem Does This Paper Address?

Osteoporosis is a major health concern, characterized by reduced bone density and increased fracture risk. Current screening methods, like the FRAX tool and bone mineral density (BMD) scans, are effective but can be costly and may lead to over-screening of individuals who are at low risk. While genome-wide association studies (GWAS) have identified genetic factors linked to osteoporosis, previous attempts to use this information for risk prediction (e.g., polygenic risk scores) haven't significantly improved upon existing clinical risk factors. The challenge is to leverage the growing availability and decreasing cost of genome-wide genotyping to create a more precise and efficient way to identify individuals at high risk for osteoporotic fractures.

Methods

The researchers employed a machine learning approach to develop a "genomically-predicted SOS" (gSOS) score. They trained six different regularized linear machine learning models (specifically L1-penalized LASSO regression) on a large dataset of individuals with genome-wide genotype data and quantitative ultrasound speed of sound (SOS) measurements. SOS is a measure that predicts fracture risk. The best-performing model was selected and then validated on an independent test dataset. This gSOS score was then integrated into the existing FRAX risk assessment tool to create a "gSOS-FRAX" score. Finally, they explored how this gSOS score could be used for pre-screening within the UK National Osteoporosis Guideline Group (NOGG) framework.

Datasets

The primary dataset used was the UK Biobank, a large-scale health resource.

  • Training dataset: 341,449 individuals with White British ancestry, measured SOS, and genotyping information.
  • Model Selection dataset: 5,335 independent individuals (1.25% of the UK Biobank cohort) used to optimize the machine learning model.
  • Test dataset: 80,027 individuals (18.75% of the UK Biobank cohort), including those with BMD measurements, used for validation and exploring screening applications.
  • A smaller cohort of non-White British participants was also used to assess the generalizability of the gSOS score.

No new benchmarks were introduced; the study aimed to improve upon existing methods like polygenic risk scores and clinical risk factors.

Key Insights and Ideas

  • Genetics can predict fracture risk beyond traditional factors: The study demonstrates that genetic information, when processed by machine learning, can capture a significant amount of variance in fracture risk that is independent of established clinical risk factors.
  • Machine learning outperforms traditional polygenic risk scores: The machine learning algorithm achieved substantially higher variance explained in SOS compared to traditional polygenic risk scores, highlighting the power of these methods for complex genetic prediction.
  • gSOS is a robust predictor, even independent of clinical factors: The gSOS score showed strong associations with incident fractures, and importantly, these associations remained significant even after accounting for clinical risk factors, suggesting it captures a distinct aspect of risk.
  • Genomic pre-screening can dramatically improve efficiency: Integrating gSOS into screening guidelines has the potential to significantly reduce the number of individuals requiring clinical visits and BMD tests, while only minimally impacting the sensitivity of identifying those who need treatment.

Key Results

  • The gSOS score explained 4.8 times more variance in SOS than FRAX clinical risk factors alone (23% vs. 4.8%).
  • A one standard deviation decrease in gSOS was associated with a 1.91-fold increased odds of incident major osteoporotic fracture, which was higher than that for measured SOS (1.60-fold) and femoral neck BMD (1.53-fold).
  • Individuals in the bottom decile of gSOS had a 3.25-fold increased risk of major osteoporotic fracture compared to the top decile.
  • A gSOS-based FRAX score identified high-risk individuals better than the CRF-FRAX score.
  • Implementing gSOS pre-screening in a simulated scenario reduced the number of required clinical visits by approximately 50% (from 2,455 to 1,273) and BMD tests by over 50% (from 1,013 to 473), with a slight decrease in sensitivity for treatment eligibility (from 99.1% to 94.7%).

Limitations and Future Work

  • Probabilistic nature: The paper emphasizes that gSOS provides a probability of fracture risk, not a deterministic diagnosis.
  • BMD as a more natural clinical measure: While gSOS performs well, BMD is a more commonly used clinical measure for fracture risk. The authors anticipate similar methods could be applied to BMD genomic prediction.
  • Missing parental hip fracture history: The study lacked data on parental hip fracture history, a known risk factor, which could potentially attenuate the clinical relevance of gSOS. This needs to be empirically tested.
  • External validation needed: The results have been internally validated but require external validation in different populations to confirm generalizability.
  • Cost-effectiveness: While efficiency gains are demonstrated, the cost-effectiveness of implementing genomic prediction in clinical practice requires further investigation.
  • Future work: The authors suggest optimizing gSOS-FRAX to include competing risks and interaction effects, and exploring its application to other heritable traits and diseases. They also highlight the potential for genomic prediction to improve healthcare delivery systems, especially as genome-wide genotyping becomes more integrated into electronic health records.

Summary updated

Abstract

Background Genomics-based prediction could be useful since genome-wide genotyping costs less than many clinical tests. We tested whether machine learning methods could provide a clinically-relevant genomic prediction of quantitative ultrasound speed of sound (SOS)—a risk factor for osteoporotic fracture. Methods We used 341,449 individuals from UK Biobank with SOS measures to develop genomically-predicted SOS (gSOS) using machine learning algorithms. We selected the optimal algorithm in 5,335 independent individuals and then validated it and its ability to predict incident fracture in an independent test dataset (N = 80,027). Finally, we explored whether genomic pre-screening could complement a UK-based osteoporosis screening strategy, based on the validated tool FRAX. Results gSOS explained 4.8-fold more variance in SOS than FRAX clinical risk factors (CRF) alone ( r 2 = 23% vs. 4.8%). A standard deviation decrease in gSOS, adjusting for the CRF-FRAX score was associated with a higher increased odds of incident major osteoporotic fracture (1,491 cases / 78,536 controls, OR = 1.91 [1.70-2.14], P = 10 -28 ) than that for measured SOS (OR = 1.60 [1.50-1.69], P = 10 -52 ) and femoral neck bone mineral density (147 cases / 4,594 controls, OR = 1.53 [1.27-1.83], P = 10 -6 ). Individuals in the bottom decile of the gSOS distribution had a 3.25-fold increased risk of major osteoporotic fracture (P = 10 -18 ) compared to the top decile. A gSOS-based FRAX score, identified individuals at high risk for incident major osteoporotic fractures better than the CRF-FRAX score (P = 10 -14 ). Introducing a genomic pre-screening step into osteoporosis screening in 4,741 individuals reduced the number of required clinical visits from 2,455 to 1,273 and the number of BMD tests from 1,013 to 473, while only reducing the sensitivity to identify individuals eligible for therapy from 99% to 95%. Interpretation The use of genotypes in a machine learning algorithm resulted in a clinically-relevant prediction of SOS and fracture, with potential to impact healthcare resource utilization. Research in Context Evidence Before this Study Genome-wide association studies have identified many loci associated with risk of clinically-relevant fracture risk factors, such as SOS. Yet, it is unclear if such information can be leveraged to identify those at risk for disease outcomes, such as osteoporotic fractures. Most previous attempts to predict disease risk from genotypes have used polygenic risk scores, which may not be optimal for genomic-prediction. Despite these obstacles, genomic-prediction could enable screening programs to be more efficient since most people screened in a population are not determined to have a level of risk that would prompt a change in clinical care. Genomic pre-screening could help identify individuals whose risk of disease is low enough that they are unlikely to benefit from screening. Added Value of this Study Using a large dataset of 426,811 individuals we trained and tested a machine learning algorithm to genomically-predict SOS. This metric, gSOS, had performance characteristics for predicting fracture risk that were similar to measured SOS and femoral neck BMD. Implementing a gSOS-based pre-screening step into the UK-based osteoporosis treatment guidelines reduced the number of individuals who would require screening clinical visits and skeletal testing by approximately 50%, while having little impact on the sensitivity to identify individuals at high risk for osteoporotic fracture. Implications of all of the Available Evidence Clinically-relevant genomic prediction of heritable traits is feasible using the machine learning algorithm presented here in large sample sizes. Genome-wide genotyping is now less expensive than many clinical tests, needs to be performed once over a lifetime and could risk stratify for multiple heritable traits and diseases years prior to disease onset, providing an opportunity for prevention. The implementation of such algorithms could improve screening efficiency, yet their cost-effectiveness will need to be ascertained in subsequent analyses.
BibTeX
@misc{forgetta2018machine,
  author = {Vincenzo Forgetta and Julyan Keller‐Baruch and Marie Forest and Audrey Durand and Sahir Bhatnagar and John P. Kemp and John A. Morris and John А. Kanis and Douglas P. Kiel and Eugène McCloskey and Fernando Rivadeneira and Helena Johannson and Nicholas C. Harvey and Cyrus Cooper and David M. Evans and Joëlle Pineau and William D. Leslie and Celia M.T. Greenwood and J. Brent Richards},
  title = {{Machine Learning to Predict Osteoporotic Fracture Risk from Genotypes}},
  year = {2018},
  doi = {10.1101/413716},
  howpublished = {bioRxiv},
}
Figure