When Healthcare Data Becomes a Challenge

Cardiovascular disease remains one of the most demanding problems in healthcare—not only because of its clinical complexity, but because timely detection can make a critical difference.

Modern healthcare systems increasingly generate large volumes of patient information through clinical databases, medical devices, and IoT-enabled sensing environments. The opportunity is enormous: if this information can be analyzed intelligently, machine learning could help identify cardiovascular risk earlier and support more informed healthcare decisions.

But there is a less visible problem.

More data does not automatically mean better decisions.

Medical datasets can contain missing values, irrelevant information, complex relationships between attributes, and variations that make accurate prediction difficult. Conventional machine-learning and deep-learning approaches can also become computationally expensive as the data becomes larger and more complex. These challenges motivated our work on a more integrated approach to cardiovascular disease detection.

The Research Question That Drove Our Work

Our central question was simple:

Can we design an intelligent diagnostic framework that can simultaneously handle imperfect medical data, identify meaningful patterns, optimize its learning process, and achieve reliable cardiovascular disease detection? 

Rather than treating missing data, feature representation, classification, and optimization as isolated problems, we designed a framework that brings these components together.

This became AH-NBDT — Alpha Hunting Optimizer-enabled Neural-Backed Decision Tree.

The research combines four important ideas:

1. Optimized KNN for missing-data handling
2. Statistical feature extraction
3. Neural-Backed Decision Tree classification
4. Alpha Hunting Optimization for parameter tuning

The motivation and contribution of the framework are explicitly centered on improving early CVD detection while reducing computational complexity and optimization difficulties.

Step 1: First, We Had to Fix the Data

One of the realities of healthcare data is that it is rarely perfect.

Missing values can interfere with subsequent analysis and classification. Instead of simply replacing missing entries with arbitrary values, our approach uses an optimized K-Nearest Neighbour (KNN) strategy.

The basic principle is intuitive: patients or observations with similar characteristics can provide useful information for estimating a missing value.

The AH optimizer is incorporated to improve the selection of parameters used by KNN. In the preprocessing stage, the original data are processed to address missing values before the information is passed to subsequent feature extraction and classification stages.

This is important because the quality of an AI decision is fundamentally connected to the quality of the information entering the model.

Step 2: Turning Raw Medical Information into Meaningful Patterns

Once the data were preprocessed, the next challenge was representation.

Instead of relying only on the original medical attributes, our framework extracts statistical characteristics from the processed data.

These include:

• Mean
• Variance
• Standard deviation
• Median
• Skewness
• Kurtosis
• Entropy
• Minimum
• Maximum
• Sum
• S-Min
• S-Max

Together, these statistical descriptors create an expanded feature representation that can capture important characteristics of the underlying data distribution. The paper describes the resulting feature vector as combining the preprocessed data with the extracted statistical features.

The idea is powerful in its simplicity:

Instead of asking the model to discover everything from raw information, we give it a richer description of the patterns hidden within that information.

Step 3: Why a Neural Network Was Not Enough

Deep neural networks are powerful, but healthcare is not only about prediction.

It is also about understanding decisions.

This is where the Neural-Backed Decision Tree (NBDT) becomes particularly interesting.

Our architecture combines the representational capability of a neural network with the structured decision-making concept of a decision tree.

The proposed architecture contains multiple dense and activation layers, with the final classification process incorporating a decision-tree structure. The NBDT framework uses hierarchical decision rules and path probabilities to reach the final prediction.

This creates an important research direction:

Can we make an AI model not only powerful, but also more structured in how it arrives at a decision?

For medical AI, this distinction matters.

A diagnostic-support system should ideally provide useful predictions while making the decision process more interpretable to researchers and healthcare professionals.

Step 4: Teaching the Model How to Search for a Better Solution

Even a well-designed neural architecture depends heavily on parameter selection.

This is where we introduced the Alpha Hunting Optimization (AHO) algorithm.

AHO was developed by integrating strategies inspired by:

• Lion Optimization Algorithm
• Coyote Optimization Algorithm
• Grey Wolf Optimization

The objective was to improve exploration and exploitation of the search space and avoid premature convergence toward inferior solutions.

One aspect that I find particularly interesting is the way the optimization strategy distinguishes between different search behaviors.

The framework incorporates Primus, Secundus, and Tertius solution strategies during exploitation, allowing candidate solutions to interact, improve, and move toward better regions of the search space.

In simple terms:

The optimizer does not blindly search for a solution. It continuously evaluates where it is, where better solutions may exist, and how the search should adapt.

The Complete AH-NBDT Pipeline

The resulting research framework can be viewed as a sequence:

Medical/IoT Data

↓

Data Preprocessing

↓

Optimized KNN Missing-Value Imputation

↓

Statistical Feature Extraction

↓

Neural-Backed Decision Tree

↓

Alpha Hunting Optimization

↓

CVD / Non-CVD Classification

The paper's system architecture illustrates this complete workflow, from medical-data acquisition through preprocessing and feature extraction to NBDT-based classification with AHO optimization.

What Did We Actually Discover?

This is where the research became particularly interesting.

We evaluated AH-NBDT using two cardiovascular datasets:

• Heart Failure Prediction dataset
• Heart Disease Detection dataset

The heart disease detection dataset contains approximately 70,000 samples, while the heart failure prediction dataset contains 918 observations. The datasets include attributes such as age, gender, blood pressure, cholesterol, glucose, ECG-related information, exercise angina, and other cardiovascular indicators.

The data were divided into 90% training and 10% testing, allowing the model to be evaluated on unseen data.

And the results were encouraging.

The Numbers That Matter

On the heart failure prediction dataset, AH-NBDT reached:

On the heart disease detection dataset:

These are not isolated numbers: the comparative experiments evaluated AH-NBDT against CatBoost, SVM, LSTM, DCNN, hybrid SVM/DCNN approaches, LOA, COA, GWO, and BHO-SVM coupled DCNN. AH-NBDT achieved the highest reported values in the paper's comparison.

But Accuracy Alone Was Not Enough

One of the strongest aspects of this research is that we did not stop at a single accuracy figure.

We examined:

• Accuracy
• Precision
• Recall
• F1-score
• Error
• K-fold performance
• ROC characteristics
• Statistical significance
• Ablation performance
• Convergence behavior
• Loss

This broader evaluation provides a more meaningful picture of model behavior.

For example, using 10-fold evaluation on the heart-failure dataset, AH-NBDT achieved 96.40% accuracy, 96.40% F1-score, 96.67% precision, and 95.864% recall, while also reporting lower error than several comparison approaches.

The Experiment That Really Tested Our Idea

Perhaps one of the most revealing experiments was the ablation study.

We wanted to understand whether combining the optimization strategies actually mattered.

When evaluated individually:

• LOA → 92.38%
• COA → 92.56%
• GWO → 92.91%

But when their complementary strategies were integrated into AHO, the resulting framework reached 96.52% accuracy.

This suggests that the improvement was not simply about adding another optimizer name to the architecture. The combination of exploration, exploitation, territorial takeover, and adaptive search behavior contributed to the optimization process.

That result was particularly valuable from a research perspective because ablation helps answer the question: “Does each design decision actually contribute?”

What I Learned From This Research

Research rarely progresses in a straight line.

One of the lessons from this work is that improving an AI system is not necessarily about making the model deeper or more complicated.

Sometimes, the bigger opportunity lies before the classifier.

Better preprocessing can improve the information entering the model.

Better feature representation can make patterns easier to distinguish.

Better optimization can make the learning process more effective.

And a structured classification mechanism can potentially make the resulting decisions easier to interpret.

This experience reinforced an important principle for me:

In intelligent healthcare, better AI is not simply about predicting more—it is about building the entire pathway from data to decision more intelligently."

The Bigger Picture

The ultimate objective of this research is not to replace physicians.

It is to explore how computational intelligence can become a supporting layer for earlier and more efficient cardiovascular risk detection.

The paper demonstrates that the AH-NBDT framework can achieve strong classification performance across two datasets, while combining data preprocessing, feature engineering, neural-backed decision-making, and optimization into a unified approach.

For me, the most exciting part is what comes next.

Healthcare is moving toward environments where sensors, connected devices, big data, machine learning, and intelligent decision-support systems increasingly interact.

The question is therefore no longer simply:

“Can AI predict cardiovascular disease?”

The more interesting question is:

“Can we build AI systems that are accurate, efficient, adaptive, interpretable, and capable of working with the imperfect data generated by real healthcare environments?”

AH-NBDT is one step toward answering that question.