📄 medicine

Interpretable machine learning analysis of apelin-12 for major adverse cardiovascular events after primary PCI in ST-elevation myocardial infarction

In a cohort of 464 STEMI patients treated with primary PCI, interpretable machine learning models demonstrated similar predictive performance to logistic regression for major adverse cardiovascular events, while SHAP analysis and non-linear modeling identified apelin-12 changes as significant predictors, though external validation is needed to confirm their incremental clinical value.

Original authors: Qingqing Wu, Rui Yan, Ping Li, Guangyao Zhai

Published 2026-08-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qingqing Wu, Rui Yan, Ping Li, Guangyao Zhai

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: why do some people who survive a massive heart attack go on to have more trouble later, while others recover smoothly? In the world of heart medicine, doctors have long used a set of standard clues—like age, blood pressure, and the size of the heart attack—to guess the future. These clues are like a classic, reliable map. But sometimes, the map misses hidden shortcuts or tricky terrain because the relationship between a clue and the outcome isn't a straight line; it's a winding road. Enter "machine learning," a type of computer brain that is great at spotting these winding roads and complex patterns that human-made maps might miss. The big question scientists are asking is: Can this super-smart computer brain find better clues than our old maps, especially when it comes to a specific chemical in the blood called "apelin-12"? This chemical is like a tiny messenger in our body that helps blood vessels relax and heal. If we can figure out exactly how this messenger behaves after a heart attack, we might be able to predict who needs extra care, potentially saving lives and preventing future heart disasters.

In this study, a team of researchers decided to put the "smart computer" to the test against the "classic map" using data from 464 patients who had a heart attack and received immediate treatment to open their blocked arteries. They wanted to see if a fancy computer algorithm could predict "Major Adverse Cardiovascular Events" (MACE)—a scary term for bad outcomes like another heart attack, heart failure, or death—better than a standard statistical model. They also wanted to see if they could finally crack the code on how apelin-12 works. Does it act like a simple switch (more is better, less is worse), or is it more like a dimmer switch where the effect changes depending on the exact level?

The researchers fed their computer three different types of "brains" to learn from: a standard logistic regression (the classic map), a Random Forest (a group of decision trees), and XGBoost (a super-charged, high-speed decision tree). They gave them 14 different clues to work with, including the patient's age, heart size, and two specific measurements of apelin-12: the amount present when they arrived at the hospital, and the percentage change in that amount 72 hours later.

Here is the twist: The fancy computer brains did not win the race. When tested on a group of patients the computer had never seen before, the "smart" models and the "classic" map performed almost exactly the same. The computer's best score was a 0.781 out of 1.0, while the classic map scored 0.750. The difference was so small that it could easily be just random luck. The paper suggests that when you don't have a massive amount of data (like millions of patients), a super-complex computer brain doesn't necessarily see things that a simpler model misses. It's like bringing a supercomputer to solve a crossword puzzle that a regular person could finish just as fast.

However, the real magic of this paper wasn't about who won the race, but what the computer learned along the way. Because the researchers used a special tool called SHAP (which acts like a flashlight to show exactly which clues the computer used and how), they could see the hidden patterns. They found that the "change rate" of apelin-12 (how much it went up or down after the attack) and the patient's age were the two most important clues.

Most importantly, the computer confirmed that apelin-12 doesn't work in a straight line. It's not just "high is good, low is bad." The data showed a "tipping point." For the change in apelin-12, if the level didn't rise by at least about 16.5% within 72 hours, the risk of future trouble jumped up. This matched a previous guess doctors had made (which was around 20%), suggesting that this chemical really does have a "sweet spot" for recovery. However, for the initial amount of apelin-12, the tipping point the computer found was 0.43 ng/mL, which was different from an earlier guess of 0.76 ng/mL. This tells us that while the chemical is important, we still need more research to nail down the exact number that matters.

The study also checked if these apelin-12 clues actually added anything new to the prediction. When they removed the apelin-12 clues from the computer, the prediction score dropped a little bit, but the evidence wasn't strong enough to say for sure that the drop was real rather than just a fluke. It's like saying, "We think these clues help, but we need to test them on more people to be 100% certain."

So, what's the takeaway? The fancy machine learning didn't magically predict the future better than the old-school math, but it did help us see the shape of the relationship between a specific chemical and heart health. It confirmed that the body's response to a heart attack is complex and non-linear. The authors suggest that while apelin-12 looks promising as a clue for future risk, we can't use it to make life-or-death decisions just yet. We need to test these findings on new groups of patients to make sure the computer isn't just making things up. Until then, the classic map and the smart computer are running neck-and-neck, and the mystery of the perfect apelin-12 threshold remains a work in progress.

Technical Summary: Interpretable Machine Learning Analysis of Apelin-12 for MACE After Primary PCI in STEMI

Problem Statement
Despite successful primary percutaneous coronary intervention (PCI) for ST-elevation myocardial infarction (STEMI), patients remain at significant risk for Major Adverse Cardiovascular Events (MACE). Conventional risk stratification tools (e.g., TIMI, GRACE) rely on additive, linear rules that may fail to capture non-linear biomarker-risk associations or complex interactions unless explicitly specified. While Apelin-12 has been identified as a prognostic biomarker in STEMI, its role within interpretable machine learning (ML) frameworks remains unclear. Specifically, it is uncertain whether Apelin-12 contributes to individualized risk prediction when modeled with flexible algorithms and whether its association with MACE is better represented as non-linear rather than linear.

Methodology
This study is a secondary analysis of a publicly available, prospectively collected, single-center cohort of 464 STEMI patients treated with primary PCI (118 MACE events, ~25.4% prevalence). The analysis utilized a fixed-horizon binary outcome (MACE within a nominal 30-month follow-up).

Data Preprocessing: The dataset was split into a training set (70%, n=325) and a held-out test set (30%, n=139) via stratified random sampling. Missing data (notably D-dimer, which was excluded from the final feature set) were handled using multiple imputation by chained equations (MICE), with training and test sets imputed separately to prevent data leakage.
Feature Selection: From 35 candidate variables, a two-step LASSO logistic regression approach was employed. Seven variables with established prognostic relevance (including Apelin-12 and its relative change, $\Delta$ Apelin-12) were exempted from penalization. The remaining variables were screened to retain a total of 14 predictors for the primary analysis, ensuring an events-per-variable ratio of approximately 5.4.
Model Development: Three models were trained on the imputed training set:
1. Standard Multivariable Logistic Regression.
2. Random Forest (500 trees, hyperparameters tuned via cross-validation).
3. XGBoost (hyperparameters tuned via grid search and cross-validation).
Evaluation Metrics: Performance was assessed on the independent test set using Area Under the Curve (AUC) with DeLong's test for pairwise comparisons, calibration (intercept and slope), Brier score, and Decision Curve Analysis (DCA).
Interpretability and Non-linearity: The XGBoost model was selected for interpretation using SHAP (SHapley Additive exPlanations) to quantify predictor contributions. To statistically assess non-linearity, Restricted Cubic Spline (RCS) and piecewise ("hinge") logistic regressions were applied to the full cohort.
Incremental Value: The contribution of Apelin-12 variables was quantified by comparing the full 14-variable XGBoost model against a reduced 12-variable model (excluding Apelin-12 and $\Delta$ Apelin-12) using bootstrapped AUC differences.

Key Results

Discrimination: All three models demonstrated similar discrimination on the test set. XGBoost achieved the highest AUC (0.781, 95% CI 0.691–0.870), followed by Random Forest (0.775) and Logistic Regression (0.750). Pairwise DeLong tests indicated no statistically significant differences between the models (all $P \ge 0.29$ ).
Calibration: Logistic Regression showed the best calibration (slope 0.783, intercept -0.012). XGBoost exhibited the greatest deviation from ideal calibration (slope 1.761, intercept 1.137), though all models showed positive net benefit in Decision Curve Analysis across lower-to-moderate threshold probabilities.
Predictor Importance: SHAP analysis of the XGBoost model ranked $\Delta$ Apelin-12 (mean |SHAP| = 0.255) and Age (0.250) as the most influential predictors. Admission Apelin-12 ranked fifth (0.133). Higher Apelin-12 and larger $\Delta$ Apelin-12 were associated with reduced predicted log-odds of MACE.
Non-linearity: Statistical assessments confirmed non-linear associations for both Apelin-12 variables.
- $\Delta$ Apelin-12: A piecewise regression identified a breakpoint at 16.5% (approximate range 15.5–19%), which closely matched a 20% cut-point reported in a prior subgroup analysis of the same cohort.
- Apelin-12: A breakpoint was identified at 0.43 ng/mL, which did not align with the previously reported 0.76 ng/mL threshold.
Incremental Value: Removing Apelin-12 variables reduced the test-set AUC from 0.781 to 0.732. However, the bootstrapped 95% confidence interval for this difference (-0.004 to 0.102) crossed zero ( $P=0.072$ ), indicating that the incremental value of Apelin-12 is suggestive but not statistically conclusive in this dataset.

Significance and Claims
The paper claims that while flexible machine learning algorithms (Random Forest, XGBoost) did not significantly outperform standard logistic regression in terms of discrimination for this modest-sized clinical dataset, the primary contribution of the ML approach lies in interpretability. By utilizing SHAP and statistical non-linearity tests, the study characterizes the structure of the Apelin-12-MACE relationship in a way that linear regression coefficients cannot.

The authors assert that:

Non-linearity is present: The association between Apelin-12 variables and MACE is non-linear, with specific breakpoints identified for $\Delta$ Apelin-12 (16.5%) and Apelin-12 (0.43 ng/mL).
Apelin-12 is a top-tier predictor: Despite the lack of statistical significance in the incremental value test, Apelin-12 and its change rate consistently rank among the most influential features in the interpretable model.
Modest expectations for ML: The results align with broader methodological evidence suggesting that in structured clinical datasets with moderate sample sizes and event counts, gains from complex algorithms over logistic regression are often limited.

The authors explicitly state that these findings are internally validated only and do not establish clinical readiness. They emphasize that external prospective validation and a definitive assessment of incremental value are required before these models or Apelin-12 thresholds can be used for clinical decision support. The study concludes that Apelin-12 warrants further evaluation within interpretable ML-based risk models for STEMI.

Technical Summary: Interpretable Machine Learning Analysis of Apelin-12 for MACE After Primary PCI in STEMI

More like this