💻 computer science

MUNNA: A Multi-strategy, Uncertainty-driven, Nested Neuro-ensemble Algorithm for Disease Diagnosis with Closed-loop Explainability Feedback

This paper introduces MUNNA, a novel three-stage hybrid framework that integrates a GWO-SA metaheuristic for feature selection, an uncertainty-weighted stacking ensemble, and a closed-loop SHAP feedback mechanism to achieve state-of-the-art disease diagnosis performance on the Breast Cancer Wisconsin dataset while addressing limitations in local optima stagnation and static ensemble weighting.

Original authors: Md. Maruf Sharker Munna

Published 2026-07-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Md. Maruf Sharker Munna

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, but instead of a crime scene, you are looking at a patient's medical chart. This chart is filled with hundreds of tiny clues—numbers representing cell sizes, shapes, and textures. The goal is to figure out if a patient is sick or healthy. In the world of science, this is called "disease diagnosis," and for a long time, computers have been helping doctors by acting as super-fast detectives. However, these computer detectives have had a few annoying habits. Sometimes, they get stuck looking at the wrong clues, thinking they are the most important ones when they aren't. Other times, they rely on a single "expert" opinion, even though a group of experts might be smarter. Finally, they often explain their answers after the fact, like a detective writing a report after the case is closed, rather than using those explanations to help them solve the mystery better in the first place.

To fix these habits, scientists use a few clever tricks. One trick is "feature selection," which is like a detective deciding which clues to keep and which to throw away so they aren't overwhelmed by noise. Another is "ensemble learning," where instead of asking one detective, you ask a whole team of different experts (some good at math, some good at patterns) and let them vote on the answer. The third trick is "explainability," which is simply asking the computer, "Why did you make that choice?" so humans can trust it. The big question is: Can we combine all these tricks into one super-smart system that learns from its own explanations to get even better?

This is exactly what a researcher named Md. Maruf Sharker Munna set out to do in a new study called MUNNA. Think of MUNNA as a high-tech detective agency that doesn't just solve the case once, but keeps refining its methods until it's perfect. The paper proposes a three-step process to diagnose diseases, specifically tested on data about breast cancer.

Step 1: The Smart Clue Hunter
First, MUNNA has to decide which clues from the medical chart are actually useful. Imagine you have a bag of 30 different tools, but you only need a few to fix a specific problem. MUNNA uses a hybrid strategy called GWO-SA. Picture a pack of wolves hunting. The "Alpha" wolf is the leader, and the pack follows it to find food (the best set of clues). But sometimes, the Alpha gets stuck in a local trap, thinking a small bush is the best spot. To fix this, MUNNA adds a "Simulated Annealing" step. This is like giving the Alpha wolf a sudden burst of energy to jump to a completely new spot just to see if there's something better there. By combining the pack's teamwork with these sudden jumps, MUNNA finds the perfect set of 20 clues out of the original 30, throwing away the noisy, useless ones.

Step 2: The Trusty Team Vote
Once the best clues are picked, MUNNA calls in a team of five different "base learners" (experts). These are different types of computer models: a Random Forest, an XGBoost, a Support Vector Machine, a K-Nearest Neighbors, and a Multilayer Perceptron. In a normal team, everyone gets an equal vote. But MUNNA is smarter. It uses uncertainty-weighted stacking. Imagine each expert is giving a prediction, but some are very confident (like saying "I'm 99% sure") while others are shaky (like saying "Maybe 55%"). MUNNA calculates how "confused" each expert is using a math concept called Shannon entropy. If an expert is confused, MUNNA listens to them less. If they are confident, it listens more. This way, the final decision isn't just a boring average; it's a dynamic vote where the most confident expert for that specific patient gets the loudest voice.

Step 3: The Feedback Loop
Here is the most unique part. Usually, after a model makes a decision, scientists use a tool called SHAP to explain why it made that choice. But MUNNA doesn't stop there. It takes those explanations and feeds them back into the very first step. It's like a detective solving a case, writing down which clues were most important, and then using that list to hire a new team of detectives for the next case who are already biased toward those important clues. This "closed-loop" happens twice. The first time, the wolves hunt for clues. Then, the SHAP tool tells them which clues were actually the heroes. The wolves then hunt again, starting with a head start on those hero clues, making the final selection even sharper.

What Did They Find?
The researchers tested this entire system on a famous dataset containing 569 samples of breast cancer cells. The results were impressive. The full MUNNA system achieved a score called AUC of 0.998 and an F1-score of 0.976. To put that in perspective, standard single models (like just using a Random Forest) scored around 0.994, and even a standard team vote without the special tricks scored 0.995. MUNNA beat them all. It managed to do this while using only 20 of the original 30 features, proving that less is often more.

However, the paper is very honest about its limits. The authors note that MUNNA's success depends heavily on how much "computing time" it is allowed to spend. When they tested it with a smaller budget (fewer wolves and fewer rounds of hunting), the advantage over a standard team vote shrank, and in some cases, the standard team was actually slightly better on average. This suggests that MUNNA isn't a magic wand that works instantly; it's a powerful tool that needs a good amount of time and energy to shine.

The study also clarifies that while the individual parts (like the wolf algorithm or the SHAP tool) are not new inventions, the specific way they are chained together—using the explanation to guide the next search, and using confidence to weight the votes—is a novel combination. The researchers didn't claim to have solved all medical diagnosis, but they showed that this specific, closed-loop approach creates a highly accurate system for this specific type of cancer data. They even adjusted the final "yes or no" decision point (the threshold) to be 0.530 instead of the usual 0.5, ensuring the system is balanced for real-world medical risks.

In short, MUNNA is a clever, self-improving detective that learns from its own mistakes and the confidence of its team members to solve medical mysteries with incredible precision, provided it has enough time to do its homework.

Technical Summary: MUNNA Algorithm for Disease Diagnosis

Problem Statement
The paper identifies three recurring weaknesses in existing machine-learning pipelines for disease diagnosis using tabular clinical data:

Feature Selection: Heuristics often stagnate at local optima.
Ensemble Learning: Static ensembles combine base learners with fixed, globally-learned weights, failing to adapt to per-patient prediction confidence.
Explainability: Explainable AI (XAI) tools like SHAP are typically used post-hoc for audit or trust-building, disconnected from the training loop rather than used to refine the model itself.

Current literature treats feature selection, ensembling, and explainability as isolated components. This paper argues that these gaps prevent pipelines from adapting to the varying trustworthiness of models for specific patients and from using evidence about feature importance to bias future feature searches.

Methodology: The MUNNA Framework
The proposed MUNNA (Multi-strategy, Uncertainty-driven, Nested Neuro-ensemble Algorithm) is a three-stage hybrid framework designed to close the loop between these components. The pipeline operates as follows:

Stage 1: Hybrid GWO-SA Feature Selection
- Uses a wrapper approach with a Grey Wolf Optimizer (GWO) combined with Simulated Annealing (SA).
- GWO performs global exploration to find candidate feature subsets.
- SA performs single-bit local refinement on the "alpha-wolf" (best incumbent) at every iteration. This allows the algorithm to escape local optima that pure GWO might suffer from.
- The fitness function balances the mean cross-validated AUC of a Random Forest surrogate against feature parsimony (penalizing large feature sets).
Stage 2: Heterogeneous Base-Learner Pool
- Trains five diverse base learners on the selected features: Random Forest, XGBoost, RBF-SVM, k-Nearest Neighbors (kNN), and a Multilayer Perceptron (MLP).
- Out-of-fold predicted probabilities are collected via 5-fold stratified cross-validation to prevent data leakage.
Stage 3: Uncertainty-Weighted Dynamic Stacking
- Instead of a fixed-weight meta-learner, MUNNA computes a per-sample confidence weight for each base learner based on the Shannon entropy of its predicted probability ( $w = 1 - H(p)$ ).
- Lower entropy (higher confidence) results in higher weights.
- The meta-learner (Logistic Regression) is trained on an augmented input vector containing the raw probabilities, the confidence weights, and the element-wise weighted probabilities. This allows the ensemble to dynamically trust different models for different patients.
Nested SHAP Feedback Loop
- After an initial pass, SHAP (SHapley Additive exPlanations) values are computed for the Random Forest base learner.
- These feature-importance scores are projected back to the full feature space and used to bias the initial population of a second GWO-SA round.
- This creates a closed loop where explainability evidence directly influences the next round of feature selection, tightening the subset around clinically influential variables.
Adaptive Threshold Calibration
- The final decision threshold is not fixed at 0.5. Instead, it is calibrated using Youden's J statistic on out-of-fold meta-predictions to maximize the balance between sensitivity and specificity.

Key Contributions
The paper claims novelty not in the individual components (GWO, SA, Stacking, SHAP are all established techniques) but in their specific closed-loop coupling:

Uncertainty-Weighted Stacking: A dynamic scheme where base learner contributions are reweighted per sample using prediction entropy, rather than using static global coefficients.
Nested SHAP Feedback: Using SHAP importances from a trained model to bias the initialization of a subsequent feature-selection search, effectively using explainability to refine the search space.
Transparent Evaluation: The authors provide a full ablation study and a "reduced-budget" cross-validation check to demonstrate where the method succeeds and where its advantages narrow.

Experimental Results
The algorithm was evaluated on the Breast Cancer Wisconsin (Diagnostic) dataset (569 samples, 30 features).

Performance: The full MUNNA pipeline achieved a held-out test AUC of 0.998 and an F1-score of 0.976.
Comparison: This outperformed single-model baselines (e.g., Random Forest AUC 0.994, XGBoost AUC 0.994), a classical stacking ensemble (AUC 0.995), and two internal ablations.
Ablation Insights:
- The largest performance gain came from the GWO-SA feature selection step.
- Interestingly, a feature-selected classical stacking ensemble performed slightly worse than the feature-selected Random Forest alone in the held-out split. The full MUNNA pipeline recovered this loss and exceeded it, suggesting that the uncertainty-weighting and nested SHAP loop are critical for making ensembling effective in this context.
Robustness Check: Under a reduced computational budget (fewer wolves, iterations, and nested rounds), the margin of MUNNA over classical stacking narrowed, and classical stacking slightly outperformed MUNNA in mean AUC (0.9956 vs. 0.9946). This indicates MUNNA's advantage is tied to the computational budget allotted to its metaheuristic search.

Significance and Claims
The paper positions MUNNA as a testable combination of established techniques rather than a categorically unprecedented algorithm. Its significance lies in:

Closing the Loop: Moving explainability from a post-hoc audit tool to an active component that refines the feature selection process.
Dynamic Ensembling: Addressing the limitation of static weights in stacking by incorporating per-sample uncertainty quantification.
Transparency: The authors explicitly report the method's limitations, including its sensitivity to computational budget, reliance on a single dataset, and the use of a surrogate fitness function, to support informed use in future clinical-decision-support research.

The authors do not claim MUNNA is a universal solution for all medical diagnosis tasks but present it as a coherent pipeline that demonstrates how integrating feature selection, uncertainty-aware ensembling, and closed-loop explainability can yield state-of-the-art results on standard benchmarks.

Technical Summary: MUNNA Algorithm for Disease Diagnosis

More like this