Prediction of 2-Year Major Adverse Limb Event-Free Survival After Percutaneous Transluminal Angioplasty and Stenting for Lower Limb Atherosclerosis Obliterans: A Machine Learning-Based Study

Tianyue Pan, Xiaolang Jiang, Hao Liu, Yifan Liu, Weiguo Fu, Zhihui DongFrontiers in Cardiovascular Medicine 2022 · DOI

TL;DR

Researchers have developed a machine learning framework to predict long-term outcomes for patients undergoing minimally invasive surgery for "clogged" leg arteries (ASO). By comparing different models, they found that an Artificial Neural Network (ANN) significantly outperforms traditional statistical methods, offering a more precise way to identify which patients will remain free of major adverse limb events two years after treatment.

What Problem Does This Paper Address?

Lower limb Atherosclerosis Obliterans (ASO) is a condition where plaque buildup narrows arteries, leading to pain, non-healing wounds, and potential amputation. While percutaneous transluminal angioplasty (PTA) and stenting—using balloons and mesh tubes to open the vessels—are the gold standard for treatment, predicting how a patient will fare long-term remains a challenge.

Existing clinical scoring systems (like WIFI or GLASS) are excellent at assessing the severity of the disease or the immediate technical success of a surgery. However, they struggle to provide a personalized prognosis. They often fail to account for the complex interplay between a patient’s unique anatomy, systemic health, and specific lesion characteristics. This paper aims to fill that gap by using machine learning to move from "one-size-fits-all" scoring to individualized risk prediction.

Methods

The researchers compared three distinct modeling approaches to predict 2-year Major Adverse Limb Event-free Survival (MFS):

  1. Logistic Regression (LR): The traditional statistical baseline.
  2. Random Forest (RF): An ensemble learning method that uses multiple decision trees to improve predictive power.
  3. Artificial Neural Network (ANN): A multilayer perceptron model designed to capture complex, non-linear relationships between variables.

The study used a chronological split (4:1 ratio) to divide the data into training and test sets. This is a more rigorous approach than random shuffling, as it simulates how a model would perform on future patients in a real-world clinical setting. Models were evaluated based on discrimination (how well they distinguish between high-risk and low-risk patients using ROCAUC) and calibration (how closely the predicted probabilities match actual outcomes).

Datasets

The study utilized a retrospectively enrolled cohort from Zhongshan Hospital, Fudan University in Shanghai, China.

  • Size: 392 patients who received successful PTA and stenting between January 2016 and December 2017.
  • Features: Data points included demographics (age, gender), risk factors (smoking, diabetes), blood markers (creatinine, fibrinogen), and detailed imaging/intervention data (lesion length, calcification grade, stent diameter).
  • Endpoint: 2-year MFS, defined as the absence of re-intervention, major amputation, or recurrence of severe symptoms with a significant drop in the ankle-brachial index (ABI).

Key Insights and Ideas

  • The "Calcification King": Using the Random Forest model's feature importance ranking, the researchers found that serious calcification (PACSS grade 4) was the single most critical factor in predicting a poor outcome. This suggests that heavily calcified vessels may restrict stent expansion or trigger negative vessel remodeling.
  • ANN vs. Random Forest: While both ML models beat the traditional logistic regression, the ANN was the clear winner in calibration. The Random Forest model tended to be "over-confident" or poorly calibrated, whereas the ANN's predicted probabilities were visually and statistically closer to the actual observed outcomes.
  • Beyond Technical Success: The study highlights that factors like lesion location and infrapopliteal runoff (the quality of blood flow below the knee) are more predictive of long-term success than the length of the stent used.

Key Results

  • Discrimination: The ANN achieved an ROCAUC of 0.80, significantly outperforming the traditional Logistic Regression (ROCAUC 0.73, $P = 0.01$). The Random Forest model sat in the middle at 0.78.
  • Specificity: The ANN showed high specificity (0.90), meaning it was particularly good at identifying patients who were not likely to experience a major adverse event.
  • Calibration: The ANN demonstrated excellent calibration ($P = 0.73$ on the Hosmer–Lemeshow test), while the Random Forest model failed significantly ($P < 0.01$), indicating its probability estimates were unreliable for clinical decision-making.

Limitations and Future Work

The authors acknowledge several hurdles before this can be used in daily practice:

  • Single-Center Bias: The data comes from one hospital, which may limit how well the model generalizes to different populations or surgical styles.
  • Sample Size: While 392 patients is substantial for this niche, it is relatively small for deep learning, potentially leading to underfitting in the ANN.
  • Missing Variables: The study did not account for different types of stents (e.g., drug-eluting vs. bare metal), which could significantly impact long-term patency.

Future work will focus on expanding the database to include multi-center data and incorporating more granular details about the mechanical properties of the stents used.

Summary updated

Abstract

Machine learning models could accurately predict 2-year MFS after PTA and stenting for lower limb ASO, in which the ANN model had better discrimination and calibration. Machine learning-derived prediction tools might be clinically useful to automatically identify candidates for PTA and stenting.
BibTeX
@inproceedings{pan2022prediction,
  author = {Tianyue Pan and Xiaolang Jiang and Hao Liu and Yifan Liu and Weiguo Fu and Zhihui Dong},
  title = {{Prediction of 2-Year Major Adverse Limb Event-Free Survival After Percutaneous Transluminal Angioplasty and Stenting for Lower Limb Atherosclerosis Obliterans: A Machine Learning-Based Study}},
  booktitle = {Frontiers in Cardiovascular Medicine},
  year = {2022},
  doi = {10.3389/fcvm.2022.783336},
}
Figure