Prediction of Major Clinical Endpoints in Atrial Fibrillation at Primary Care Level using Longitudinal Learning Stances

This study develops superior longitudinal machine learning models that outperform traditional clinical scores in predicting six major adverse clinical endpoints for atrial fibrillation patients within a Portuguese primary care cohort, while also identifying key risk determinants and introducing a prototype decision-support tool.

The Gist

Imagine your heart is like a busy city intersection. Atrial Fibrillation (AF) is when the traffic lights start flashing randomly, causing cars (blood) to swirl instead of flowing smoothly. This chaos creates a high risk of accidents: traffic jams (heart failure), crashes (strokes), or the city shutting down entirely (death).

For years, doctors have tried to predict these accidents using a "Static Scorecard." Think of this like a weather forecast based only on today's temperature. They ask: "Are you old? Do you have high blood pressure? Did you have a stroke before?" They add up points (like the famous CHA2DS2-VASc score) to guess your risk.

The Problem: Life isn't static. Your health changes every day. A scorecard that only looks at a single snapshot misses the story. It doesn't know if your blood pressure has been slowly creeping up for months, or if your kidney function has been wobbling like a shaky table.

The Solution: This paper introduces a "Time-Traveling Detective" (Machine Learning). Instead of just looking at a photo, the detective watches a 25-year movie of the patient's life using their medical records.

Here is how the researchers built this detective, broken down simply:

1. The Data: A Massive Library of Medical Diaries

The team gathered the electronic health records of over 7,200 patients from a hospital in Portugal. They didn't just look at one visit; they looked at the entire history of every patient over decades.

• The Analogy: Imagine trying to predict who will win a marathon. A traditional method looks at who is wearing the fastest shoes right now. This new method looks at the runner's entire training log, their diet changes, their sleep patterns, and how their speed has changed over the last 10 years.

2. The Method: Teaching the AI to "Watch the Clock"

The researchers taught computer algorithms (like XGBoost and Random Forest) to understand time.

• Static View: "The patient has high cholesterol."
• Longitudinal View (The New Way): "The patient's cholesterol was normal 5 years ago, spiked last year, and has been slowly rising for the last 6 months."
• The Magic: By feeding the AI this "movie" of data, it could spot subtle patterns that human doctors or simple scorecards miss. It learned that changes in health are often more dangerous than the health status itself.

3. The Results: Outsmarting the Old Scorecards

The team tested their new AI against the old "Static Scorecards" (CHA2DS2-VASc and GARFIELD-AF) to see who could better predict six scary outcomes: Stroke, Death, Heart Failure, Hospital Visits, and more.

• Predicting Death: The old scorecard (GARFIELD-AF) was decent, getting a score of 0.72 (on a scale where 1.0 is perfect). The new AI, using the "time-travel" method, scored 0.78. It was significantly better at spotting who was in danger.
• Predicting Stroke: The old scorecard (CHA2DS2-VASc) scored 0.59 (barely better than flipping a coin). The new AI scored 0.65. It wasn't perfect, but it was a clear step up.

The "Aha!" Moments:
The AI found some surprising clues that the old scorecards ignored:

• The "Obesity Paradox": Surprisingly, patients with lower body weight were often at higher risk. The AI figured this out because in sick, elderly populations, low weight often means the body is wasting away (frailty), not that they are healthy.
• The "Height" Clue: Taller people seemed to have a slightly lower risk. This is a weird finding that the AI picked up, suggesting our height might be linked to how our hearts developed early in life.
• The "Medication" Clue: The AI realized that taking certain drugs (like insulin or diuretics) was a stronger warning sign than the disease itself. It meant the patient was already being treated for something serious.

4. The Tool: A Digital Co-Pilot for Doctors

The researchers didn't just stop at the math. They built a prototype app (a Decision Support Tool).

• How it works: A doctor in a primary care clinic can type in a patient's data. The app instantly analyzes the patient's history, compares them to the 7,200 people in the study, and says: "Based on how this patient's health has changed over time, they have a high risk of a heart event in the next 6 months."
• The Benefit: It gives the doctor a "second opinion" that considers the full timeline of the patient's life, not just the symptoms they have today.

The Bottom Line

This paper is like upgrading from a paper map to a GPS with live traffic.

• Old Way: "You are in a city known for traffic; be careful." (Static Score)
• New Way: "You are in a city known for traffic, and your car has been making a weird noise for three weeks, and you took a wrong turn yesterday. You are likely to crash soon. Let's reroute." (Longitudinal AI)

While the AI isn't perfect yet (it needs more data and testing), it proves that looking at the story of a patient's life, rather than just a single snapshot, can save lives by predicting heart attacks and strokes before they happen.

Technical Summary

1. Problem Statement

Atrial Fibrillation (AF) is the most prevalent cardiac arrhythmia globally, significantly increasing the risks of stroke, heart failure, and mortality. Current approaches to risk stratification face two primary limitations:

• Static Nature: Traditional risk scores (e.g., CHA₂DS₂-VASc, HAS-BLED) and existing Machine Learning (ML) models often rely on point-in-time data, failing to capture the dynamic, evolving nature of patient physiology and behaviors over time.
• Data Underutilization: State-of-the-art predictors frequently neglect the integration of longitudinal data (temporal trends) and specific data modalities (electrophysiological signals, medication history, risk behaviors), limiting their ability to model individual risk progression across different prognostic windows.

The study aims to develop superior ML models tailored to predict adverse clinical events in an AF cohort by leveraging longitudinal Electronic Health Records (EHRs) and to provide a clinical decision-support tool.

2. Methodology

Dataset

• Source: Anonymized EHRs from the Unidade Local de Saúde de Matosinhos (ULSM), Portugal.
• Cohort: 7,203 patients diagnosed with AF between 2012 and 2021 (aged >40).
• Features: 167 variables including demographics, clinical features (lab results, pharmaceuticals, surgical acts), and temporal context.
• Target Endpoints: Six clinical outcomes were predicted:
  1. Stroke/Systemic Embolism (SE)
  2. All-cause Death
  3. Cardiovascular (CV) Death
  4. Heart Failure (HF) Hospitalizations
  5. Inpatient Visits
  6. Acute Coronary Syndrome (ACS)

Data Preprocessing

• Cleaning: Recalculation of time variables relative to AF diagnosis, outlier capping/log-transformation, and correction of implausible values.
• Feature Engineering:
  • Creation of three dataset versions: Static (snapshot), Slope-based (rate of change), and Longitudinal (full temporal history).
  • Engineering of binary indicators for vascular disease, diabetes, kidney function, and medication use.
  • Handling missing data via targeted removal and mean imputation.
• Imbalance Handling: Addressed class imbalance using random sampling and SMOTE (Synthetic Minority Over-sampling Technique).

Models & Evaluation

• Algorithms: Implemented Naïve Bayes, Logistic Regression (LR), Decision Tree, Random Forest (RF), XGBoost, and Multi-Layer Perceptron (MLP).
• Baselines: Compared against classical risk calculators: CHA₂DS₂-VASc (for stroke) and GARFIELD-AF (for stroke and mortality).
• Training Strategy:
  • 5-fold cross-validation.
  • Hyperparameter optimization via Bayesian optimization using the F2 score (prioritizing recall/sensitivity to minimize false negatives).
  • Model selection based on rank fusion of F1, F2, and AUC.
• Explainability: SHAP (SHapley Additive exPlanations) values were used to interpret feature importance.

3. Key Contributions

1. Longitudinal Modeling: Demonstrated that incorporating temporal features (time-aware models) significantly outperforms static models and traditional risk scores in predicting AF-related outcomes.
2. Systematic Comparison: Provided a rigorous benchmark comparing ML predictors against established clinical scores (CHA₂DS₂-VASc, GARFIELD-AF) across multiple endpoints.
3. Explainable AI: Identified specific determinants of risk (e.g., temporal variations in creatinine and glycemia) that traditional scores miss, offering new clinical insights.
4. Clinical Decision Support Tool (CDST): Developed a prototype tool (FastAPI backend + Dash frontend) allowing clinicians to input patient data and visualize real-time predictions, bridging the gap between research and primary care practice.

4. Key Results

All-Cause Mortality (6-month horizon)

• Performance: The Longitudinal XGBoost model achieved an AUC of 0.779, significantly outperforming the GARFIELD-AF baseline (AUC 0.715) and static models.
• Key Findings:
  • Age was the strongest predictor.
  • Temporal features (changes in creatinine, glycemia, and HDL over time) provided predictive value beyond single measurements.
  • Obesity Paradox: Lower BMI was associated with higher risk, likely due to frailty.
  • Lipid Patterns: Lower total cholesterol and LDL were linked to higher risk, likely reflecting widespread statin use in the cohort.

Stroke/Systemic Embolism (1-year horizon)

• Performance: The Slope-based Logistic Regression model achieved an AUC of 0.651, outperforming both CHA₂DS₂-VASc (AUC 0.588) and GARFIELD-AF (AUC 0.633).
• Key Findings:
  • Incorporating slope and longitudinal features improved performance over static models.
  • Unconventional predictors like height showed an inverse relationship with risk, suggesting a link to early-life cardiovascular health factors.
  • Most patients in the cohort were already classified as "high risk" by CHA₂DS₂-VASc, limiting its discriminative power, whereas ML models could better stratify risk within this high-risk group.

General Observations

• Time Horizons: Model performance generally improved as the prediction horizon increased (e.g., 3-month to 1-year), though the highest AUC for mortality was observed at 3 months.
• Feature Importance: Traditional risk factors like hypertension often lost predictive value due to their high prevalence (81.8%) in the cohort. Instead, medication usage (insulin, digoxin, loop diuretics) and specific lab trends became more significant predictors.

5. Significance and Future Work

• Clinical Impact: The study proves that time-aware ML models can provide more accurate prognostic assessments than current standard-of-care scores, potentially enabling earlier interventions for high-risk AF patients in primary care.
• Implementation: The prototype CDST demonstrates the feasibility of integrating complex ML models into user-friendly clinical workflows.
• Limitations: The study is limited by its single-center cohort (Portugal), which may affect generalizability. Missing data on AF type, pulse, dementia, and specific anticoagulant classes (NOAC vs. VKA) constrained the models.
• Future Directions:
  • External validation on independent cohorts.
  • Integration of raw ECG signals and cardiac imaging.
  • Development of deep learning architectures to better handle irregular sampling and complex temporal relationships.
  • Refinement of the CDST through iterative clinician feedback.

In conclusion, this work establishes a foundation for moving AF risk prediction from static, point-based scoring to dynamic, longitudinal learning systems, offering a more nuanced and accurate approach to managing atrial fibrillation complications.