Using multiomic data to predict postoperative complications after major surgery in the UK Biobank cohort

In a large UK Biobank cohort study, adding high-dimensional metabolomic and proteomic data to standard clinical variables failed to improve the prediction of postoperative complications, likely due to the temporal gap between sample collection and surgery, although transfer learning techniques enhanced model stability.

The Gist

The Big Question: Can We Predict Surgery Risks with a "Biological Snapshot"?

Imagine you are about to go on a very long, difficult hike (major surgery). You want to know: Will I get injured on the trail?

Traditionally, doctors look at your "hiker profile": your age, how much weight you carry, if you have a bad knee, or if you've hiked before. These are your clinical variables. They are good, but they don't tell the whole story.

Recently, scientists have started using high-tech tools to take a "biological snapshot" of your body. This includes looking at your metabolites (tiny chemicals your body makes) and proteins (the building blocks of your cells). The hope was that if we looked at these tiny details, we could predict surgery complications much better than just looking at your age and medical history.

The Study:
Researchers in the UK took data from 158,000 people who had major surgery. They tried to build a computer model to predict six scary complications: heart rhythm issues, kidney failure, heart attacks, confusion (delirium), strokes, and infections.

They built two types of models:

1. The "Old School" Model: Used only age, sex, and medical history.
2. The "Super-Model": Used age, history, plus the high-tech biological snapshots.

The Twist: The Snapshot Was Taken Too Long Ago

Here is the catch: The biological snapshots (blood tests) were taken from these people an average of 6 years before they had their surgery.

Think of it like this:

• You want to predict if a car will break down during a race next week.
• You have a detailed photo of the engine taken six years ago.
• The photo shows the engine was clean and new back then.
• But today, the engine might be rusty, or the oil might be dirty, or a new part might be failing.

The researchers found that the "Super-Model" (using the 6-year-old biological data) did not work any better than the "Old School" model.

What They Found (The Results)

1. The Basics Work Great: The models using just age and medical history were actually quite good at predicting who would get sick after surgery. They were like a reliable, old map that gets you to the destination 80% of the time.
2. The Fancy Data Didn't Help: Adding the complex biological data (the 6-year-old snapshots) didn't make the prediction any more accurate. It was like adding a GPS to a car that already has a working compass; the GPS was just showing where the car was six years ago, not where it is right now.
3. The "Time Gap" Problem: The main reason the fancy data failed is that biology changes fast. A chemical level that was normal six years ago might be totally different the day before surgery. The "snapshot" was too blurry and too old to be useful for predicting an immediate event.

The "Transfer Learning" Experiment

The researchers also tried a clever trick called Transfer Learning.

• The Idea: Since there were very few people with surgery complications in the data, they tried to learn from a bigger group of people who had similar health issues but didn't have surgery.
• The Analogy: Imagine trying to learn how to drive a race car (surgery complications) but you've never driven one. So, you practice on a regular sedan (non-surgical health data) first, hoping the skills transfer over.
• The Result: This helped the computer models become more stable and less shaky, but it still didn't beat the simple "Old School" model. It proved that the biology of chronic illness and surgery complications are related, but it didn't solve the "time gap" problem.

The Bottom Line

Don't throw away the old map yet.

The study concludes that for now, standard clinical checks (age, weight, medical history) are still the best way to predict surgery risks. Adding expensive, high-tech blood tests taken years in advance doesn't give doctors any extra useful information.

The Takeaway for the Future:
If scientists want to use these fancy biological tools, they need to take the "snapshot" right before the surgery (like checking the engine today before the race), not six years ago. Only then might they see a real benefit.

In short: We have powerful new tools, but we were trying to use yesterday's weather report to predict a storm happening tomorrow. It just doesn't work. Stick to the basics for now!

Technical Summary

1. Problem Statement

Postoperative complications (e.g., atrial fibrillation, acute kidney injury, delirium) significantly impact patient morbidity, mortality, and healthcare resource utilization. Current pre-operative risk stratification relies on demographic factors and clinical scoring systems (e.g., ASA grade), which fail to capture the biological heterogeneity of individual patients.

While high-dimensional molecular assays (metabolomics and proteomics) offer a granular view of physiological states, their utility in the peri-operative setting remains unproven. Key challenges include:

• Temporal Resolution: Biobank samples are often collected years before the surgical event, potentially missing acute physiological changes.
• Data Scarcity: Rare postoperative complications result in small sample sizes relative to the high dimensionality of omic data (the "curse of dimensionality").
• Incremental Value: It is unclear if adding omic data to standard clinical variables provides clinically meaningful improvements in predictive accuracy.

2. Methodology

Study Design & Cohort

• Data Source: UK Biobank (n = 158,156 individuals undergoing major inpatient surgery).
• Omic Subsets: A subset of participants had metabolomic (Nightingale NMR panel, ~251 biomarkers) and/or proteomic (Olink Explore platform, up to 2,922 biomarkers) data.
• Outcomes: Six primary postoperative complications defined by ICD-10 codes within 30 days of surgery:
  1. Atrial Fibrillation (AF)
  2. Acute Kidney Injury (AKI)
  3. Acute Myocardial Infarction (AMI)
  4. Delirium
  5. Stroke
  6. Surgical Site Infection (SSI)
• Temporal Gap: The median time between omic sampling and surgery was 6 years (IQR 2.9–9.4).

Modeling Approaches
The study employed two primary modeling strategies:

• A. Direct Postoperative Modelling:

  • Baseline Models: Elastic Net logistic regression using standard clinical covariates (Age, Sex, Charlson Comorbidity Index, admission method, operation category).
  • Omic Models: Baseline variables + Metabolomic, Proteomic, or Multiomic features.
  • Dimensionality Reduction: Multiomic data was processed using Autoencoders (AEs) to compress inputs into a 16-dimensional latent space. Both early integration (combined AE) and late integration (separate AEs) were explored.
  • Handling Imbalance: Models were trained within a Balanced Bagging Classifier (50 base estimators) to address class imbalance (1:1 case-control ratio via undersampling).
  • Validation: Adaptive repeated stratified k-fold cross-validation (20 repeats of 3-fold for N<100; 12 repeats of 5-fold for N≥100).

• B. Transfer Learning:

  • Source Domain: Non-surgical UK Biobank participants with omic data.
  • Target Domain: Surgical cohort.
  • Process: Models were first trained on non-postoperative equivalents of the outcomes (e.g., chronic AF, non-surgical stroke) using clinical + omic data. The predicted probabilities from these source models were then used as a single feature in a parsimonious logistic regression model to predict the postoperative outcome.

Evaluation Metrics

• Primary: Area Under the Receiver Operating Characteristic Curve (AUROC), Sensitivity (Recall), and Specificity.
• Statistical Significance: Corrected Resampled t-test with Benjamini-Hochberg False Discovery Rate (FDR) correction.

3. Key Results

Sample Sizes

• Metabolomics: 144 (AMI) to 1,596 (AKI) cases.
• Proteomics: 27 (AMI) to 289 (AKI) cases.
• Multiomics: 15 (AMI) to 219 (AKI) cases.
• Note: Many proteomic and multiomic subsets fell below recommended sample sizes for robust predictive modeling.

Performance Findings

1. Baseline Clinical Models: Achieved robust performance across all phenotypes (AUROC 0.72–0.88; Sensitivity 0.71–0.80).
2. Impact of Omic Data:
   • Adding metabolomic or proteomic features to baseline clinical models did not produce clinically meaningful improvements in AUROC, sensitivity, or specificity for any outcome.
   • For SSI, proteomic features improved specificity slightly but did not improve overall discrimination.
   • For AMI and Stroke, sample sizes were too small to draw definitive conclusions regarding omic utility, though trends were negative.
3. Transfer Learning:
   • Transfer learning from non-postoperative domains improved model stability (reduced variance) compared to direct modeling, particularly for rare outcomes like AMI and Stroke.
   • However, transfer models did not outperform the baseline clinical models in terms of predictive accuracy.
   • For SSI, using an alternative source domain (skin infections) resulted in reduced sensitivity compared to baseline models, highlighting the risk of mismatched source domains.

4. Key Contributions

• Large-Scale Validation: This is one of the largest studies to evaluate multiomic data for postoperative risk stratification, utilizing a diverse cohort of over 150,000 surgical patients.
• Methodological Rigor: The study employed advanced machine learning techniques (Elastic Net, Autoencoders, Balanced Bagging) and rigorous cross-validation to prevent overfitting and data leakage.
• Transfer Learning Application: It successfully demonstrated that transfer learning can stabilize models for rare surgical complications by leveraging larger non-surgical datasets, though it did not solve the fundamental lack of predictive signal from distant omic data.
• Temporal Insight: The study explicitly quantified the impact of the temporal gap (median 6 years) between biomarker collection and the surgical event, identifying this as a critical limitation.

5. Significance and Conclusion

Clinical Implications

• Parsimony Prevails: The study concludes that standard clinical variables remain the most robust and cost-effective tools for pre-operative risk stratification.
• No Current Justification for Omics: Collecting high-throughput metabolomic or proteomic data at a temporal distance from surgery (e.g., during routine biobank enrollment) does not justify the cost and logistical complexity, as it adds no incremental predictive value.
• Future Directions:
  • Timing is Critical: Omic data may only be useful if collected in the immediate peri-operative period to capture acute physiological decline and inflammatory responses.
  • Transfer Learning Potential: The success of transfer learning suggests shared biological architectures between chronic and acute phenotypes, warranting further research with larger non-surgical cohorts to refine peri-operative risk models.

Limitations

• Healthy Volunteer Bias: UK Biobank participants are generally healthier than the general population.
• Retrospective Outcomes: Complications were derived from clinical coding, potentially leading to underreporting.
• Missing Variables: Key clinical predictors (e.g., ASA grade, anesthesia type) were unavailable in the linked data.
• Sample Size Constraints: Despite the large cohort, the specific subset with omic data and rare complications was often underpowered for high-dimensional modeling.

Final Verdict: The addition of metabolomic and proteomic data collected years prior to surgery does not improve the prediction of postoperative complications beyond standard clinical variables. The "snapshot" nature of biobank omics fails to capture the acute physiological state required for peri-operative risk assessment.