Federated Learning with Multi-Partner OneFlorida+ Consortium Data for Predicting Major Postoperative Complications

This study demonstrates that federated learning models trained on a large, multicenter dataset from the OneFlorida+ Consortium effectively predict major postoperative complications and mortality with robust generalizability and privacy preservation, achieving performance comparable to or superior to both local and centralized models.

The Gist

Imagine you are trying to teach a robot how to predict which patients will get sick after surgery. To do this well, the robot needs to learn from thousands of past cases.

The Problem: The "Secret Recipe" Dilemma
In the real world, hospitals are like different chefs in different cities. Each chef (hospital) has their own secret recipe book (patient data) filled with unique stories about their patients.

• The Old Way (Local Learning): If Chef A only reads their own book, they become great at predicting outcomes for their specific neighborhood, but they might fail miserably when a patient from a different city comes in.
• The Dangerous Way (Central Learning): If all chefs dump their secret recipe books into one giant pot in the middle of the room, the robot learns from everyone. But this is risky! It's like handing your private diary to a stranger. Hospitals can't do this because of strict privacy laws (HIPAA) and patient trust.

The Solution: The "Group Chat" (Federated Learning)
This paper introduces a clever solution called Federated Learning. Think of it as a secure group chat where the chefs can share their lessons learned without ever sharing their actual recipe books.

1. The Setup: Instead of moving the data, the robot (the AI model) travels to each hospital.
2. The Training: The robot learns from the local data at Hospital A, then goes to Hospital B, then Hospital C.
3. The Handshake: When the robot leaves a hospital, it doesn't take any patient names or records. It only takes a "summary of what it learned" (mathematical updates).
4. The Mastermind: A central server collects these summaries from all hospitals, mixes them together to create a "Super Robot," and sends the improved version back out.

What They Did
The researchers used this method with five different hospitals in Florida (the OneFlorida+ Consortium). They had data on nearly 360,000 patients who underwent major surgeries between 2012 and 2023.

They asked the Super Robot to predict four scary things:

• Will the patient need to go to the ICU?
• Will they need a breathing machine (ventilator)?
• Will they develop kidney failure (AKI)?
• Will they die in the hospital?

The Results: The Best of Both Worlds
The researchers tested three types of robots:

1. The Local Robot: Trained only on one hospital's data. (Good for that hospital, bad for others).
2. The Central Robot: Trained on all data mixed together. (Great at predicting, but required breaking privacy rules).
3. The Federated Robot (SCAFFOLD): The "Group Chat" robot.

The Winner: The Federated Robot won!

• It was just as smart as the "Central Robot" (which had all the data).
• It was much smarter than the "Local Robots" when visiting new hospitals.
• Crucially: No patient data ever left the hospital walls. Privacy was 100% preserved.

A Special Trick: The "Surgeon's Signature"
The researchers also tried a "fine-tuning" trick. They realized that just like a student learns better with a specific teacher, a model might work better if it knows which surgeon is operating.
They added the surgeon's identity as a special "personalized feature" to the model. This was like giving the robot a little note saying, "Dr. Smith tends to be very careful with older patients." This made the predictions slightly even more accurate for that specific hospital.

Why This Matters
This study is a big deal because it proves we can build super-smart medical AI without violating patient privacy.

• For Doctors: It's like having a crystal ball that works in any hospital, not just the one where it was built. It can warn them, "Hey, this patient looks high-risk for kidney trouble; let's prepare extra fluids now."
• For Patients: It means better care and fewer surprises, all while keeping their medical history safe and sound.

In a Nutshell
This paper shows that we don't have to choose between privacy and progress. By using Federated Learning, we can teach AI to be a world-class surgeon's assistant by letting hospitals "share wisdom" without ever "sharing secrets."

Technical Summary

1. Problem Statement

The study addresses the critical challenge of predicting major postoperative complications (Intensive Care Unit admission, Mechanical Ventilation, Acute Kidney Injury, and in-hospital mortality) using Electronic Health Records (EHR).

• Limitation of Current Models: Existing machine learning models are typically trained on single-center data, leading to poor generalizability across diverse patient populations and clinical settings.
• Data Privacy Barrier: While pooling data from multiple institutions (Central Learning) improves model robustness, it raises significant legal and ethical concerns regarding patient privacy and data security, often preventing the creation of large, diverse multicenter datasets.
• The Gap: There is a need for a methodology that leverages the statistical power of multicenter data without compromising patient privacy, specifically tailored for the non-IID (Non-Independently and Identically Distributed) nature of healthcare data across different hospitals.

2. Methodology

The researchers conducted a retrospective, longitudinal, multicenter cohort study using data from the OneFlorida+ Data Trust, involving five healthcare partners.

Data Cohort

• Scope: 358,644 adult patients undergoing 494,163 major inpatient surgical procedures (2012–2023).
• Partners: Five institutions (Partners 1, 2, 3, 4, 6).
  • Development Sites: Partners 3, 4, and 6 (used for training, validation, and internal testing).
  • External Validation Sites: Partners 1 and 2 (specifically, the "External Validation" cohort mentioned in results refers to data from a distinct set of patients, likely Partner 1/2 or a hold-out set, used to test generalizability).
• Features: 99 preoperative features including demographics, socioeconomic factors, comorbidities, medications, and preoperative labs.
• Outcomes: Postoperative ICU admission, Mechanical Ventilation (MV), Acute Kidney Injury (AKI), and in-hospital mortality.

Learning Paradigms Compared

The study evaluated three distinct approaches:

1. Local Learning: Models trained exclusively on data from a single center.
2. Central Learning: Models trained on a pooled dataset where raw data from all centers is aggregated.
3. Federated Learning (FL): Models trained locally at each site with model updates (weights) aggregated centrally, ensuring raw patient data never leaves the local server.

Federated Learning Algorithms

The study implemented and compared several FL algorithms tailored for neural networks and tree-based models:

• FedAvg: Standard averaging of model weights.
• FedProx: Adds a proximal term to penalize deviation from the global model, handling heterogeneity.
• SCAFFOLD: Uses control variates to correct for client drift, specifically designed for Non-IID data distributions.
• Federated XGBoost: A tree-based ensemble method implemented via Federated XGBoost.

Model Architecture

• A deep learning architecture was used, processing continuous/binary features via fully connected layers and high-cardinality features via embedding layers.
• The latent representations were merged and passed through four separate output branches to predict the four distinct outcomes simultaneously (multi-task learning).
• Sensitivity Analysis: The FL model was fine-tuned at each site by adding a personalized feature: Surgeon's Identity, to assess if local context could further improve performance.

3. Key Contributions

• Scalable Multicenter FL Implementation: Successfully developed and validated FL models across five diverse healthcare institutions, demonstrating the feasibility of FL in a large-scale surgical context.
• Algorithmic Comparison: Provided a rigorous comparison of FL algorithms (FedAvg, FedProx, SCAFFOLD, Federated XGBoost) in a real-world clinical setting, identifying SCAFFOLD as the superior algorithm for handling the Non-IID nature of surgical data.
• Privacy-Preserving Generalizability: Demonstrated that FL models can achieve performance comparable to or better than Central Learning models (which require raw data sharing) while maintaining strict data privacy.
• Hybrid Approach Validation: Showed that fine-tuning global FL models with local features (e.g., surgeon identity) yields marginal but consistent improvements, suggesting a hybrid "Global-Local" strategy is optimal.

4. Results

• Algorithm Performance:
  • SCAFFOLD consistently outperformed or matched FedAvg, FedProx, and Federated XGBoost across all outcomes and sites.
  • AUROC Scores: Ranged from 0.73 to 0.97 depending on the outcome and site.
    • Highest: Mechanical Ventilation (0.96–0.97) and In-hospital Mortality (0.96–0.97).
    • Lowest: AKI in external validation (0.73), likely due to class imbalance and data distribution shifts.
  • AUPRC Scores: Varied significantly due to class imbalance (e.g., mortality rates were low), but SCAFFOLD generally maintained the best balance between precision and recall.
• Comparison with Other Paradigms:
  • vs. Local Learning: Local models performed well on their own test data but failed significantly when applied to other sites (poor generalizability). The SCAFFOLD FL model achieved comparable or superior performance to the best local model at every site, proving its ability to generalize across diverse populations.
  • vs. Central Learning: SCAFFOLD achieved comparable or superior AUROC and AUPRC scores compared to the Central Learning model, validating that FL does not sacrifice predictive power for privacy.
• Sensitivity Analysis:
  • Fine-tuning the SCAFFOLD model with the "Surgeon's Identity" feature resulted in slight but consistent improvements in AUROC and AUPRC across most outcomes (e.g., ICU admission AUROC improved from 0.85 to 0.87 at Partner 3).

5. Significance and Implications

• Clinical Decision Support: The study validates the feasibility of deploying privacy-preserving AI in clinical workflows. These models can be used preoperatively to identify high-risk patients, allowing for tailored interventions (e.g., fluid optimization, anemia management) and better resource allocation (ICU bed planning).
• Overcoming Data Silos: The research provides a blueprint for overcoming the "data silo" problem in healthcare. It proves that institutions can collaborate to build robust models without violating HIPAA or patient trust by sharing raw data.
• Handling Non-IID Data: The success of the SCAFFOLD algorithm highlights the importance of selecting the right FL aggregation strategy for heterogeneous medical data, where patient demographics and practice patterns vary significantly between hospitals.
• Future Directions: The authors suggest that future work should focus on standardizing data cleaning across centers, addressing technical challenges like network latency, and incorporating more granular surgical data (e.g., anesthesia types, exact timing) to further refine model accuracy.

In conclusion, this study establishes that Federated Learning (specifically using the SCAFFOLD algorithm) is a robust, generalizable, and privacy-preserving solution for predicting postoperative complications, offering a viable path toward large-scale, collaborative AI in healthcare.