Enhancing healthcare diagnostics with artificial intelligence: A hybrid machine learning model for early disease detection

Related work

The application of AI in healthcare diagnostics has seen rapid growth, with numerous studies demonstrating the efficacy of ML and deep learning (DL) models in predicting and detecting chronic diseases. For instance, CNNs have been widely used for feature extraction from medical imaging and structured data, achieving dermatologist-level accuracy in skin cancer classification (Esteva et al., 2017).^[6] Similarly, random forests have proven effective in handling EHR-based predictions due to their interpretability, efficiency and ability to manage high-dimensional data (Breiman, 2001).^[3]

Hybrid models combining CNNs with other algorithms have shown promising results in addressing the complexities of chronic disease detection. Zhou et al. (2023)^[5] proposed a CNN-AdaBoost and random forest (ABRF) model integrating CNN for feature extraction from Electronic Medical Records (EMRs) and an ensemble of ABRF for classification, achieving high precision (89.28%) and recall (88.89%) on a dataset of 18,590 Chinese EMRs covering 10 chronic diseases. This model excelled in handling imbalanced data and reducing misdiagnosis for diseases with similar symptoms.

Salman and Gupta (2023)^[4] developed a hybrid framework using random forest for feature extraction and logistic regression for classification, improving prediction accuracy for chronic diseases by addressing outliers and feature selection issues. Their approach was simulated on Python, emphasising economic, social and epidemiological factors.

Suman et al. (2024)^[7] introduced a hybrid CNN-A bidirectional long short-term memory (BiLSTM) model with principal component analysis (PCA) for feature reduction, achieving accuracies up to 98.37% for heart disease, 83.76% for diabetes and 95% for kidney disease on University of California, Irvine (UCI) datasets, outperforming traditional ML algorithms by approximately 10%.

Shambharkar et al. (2023)^[8] proposed a convolutional neural network (CNN)-K-nearest neighbors (KNN) hybrid for early chronic disease detection, utilising CNN for automatic feature extraction and KNN for distance-based classification, resulting in 95% accuracy, 94% precision and 99% recall, surpassing Naïve Bayes, decision tree and logistic regression.

A comprehensive review by Sadr et al. (2025)^[9] synthesises ML and DL applications across 16 diseases, highlighting CNNs and hybrid models in enhancing diagnostic accuracy for conditions such as diabetes, cardiovascular diseases and kidney disease, while noting challenges in data quality and interpretability.

Other notable works include Chollet (2018)^[10] on DL fundamentals and Rajkomar et al. (2018)^[2] on scalable DL with EHRs. Our hybrid CNN-random forest model builds on these foundations, optimising for both accuracy and efficiency in EHR-based diagnostics.

Dataset

The study employs a dataset of 10,000 anonymised EHRs sourced from a public health repository, similar to those used in prior research (Rajkomar et al., 2018).^[2] The dataset encompasses a wide range of features, including demographic information (age and gender), vital signs (blood pressure and heart rate), laboratory results (glucose levels, cholesterol and glycated haemoglobin) and medical history indicators (previous diagnoses and medications). Preprocessing steps include handling missing values through mean imputation for numerical features and mode imputation for categorical ones, normalising numerical data using Min-Max scaling to a [0, 1] range and one-hot encoding categorical variables to prevent ordinal assumptions. This ensures that the data are clean, balanced and ready for model input, addressing common issues in EHR data such as incompleteness and variability (Zhou et al., 2023).^[5]

Model architecture

The proposed hybrid model integrates CNN for deep feature extraction and random forest for classification, drawing inspiration from successful hybrids like CNN-ABRF (Zhou et al., 2023).^[5] The CNN component consists of three convolutional layers with 32, 64 and 128 filters, respectively, each using a kernel size of 3, ReLU activation, and followed by max-pooling layers (pool size 2) to reduce spatial dimensions and computational load. A dropout layer (rate 0.5) is incorporated after each pooling to mitigate overfitting. The output features from the CNN are flattened into a 1D vector and fed into the random forest classifier, comprising 100 decision trees with a maximum depth of 10 to balance complexity and generalisation (Breiman, 2001).^[3] This architecture leverages CNN’s ability to capture local patterns in EHR data (e.g., temporal trends in vital signs) and random forest’s ensemble voting for robust, interpretable predictions.

Training process

The dataset is split into 70% training, 15% validation and 15% testing sets. The CNN is trained using the Adam optimizer with a learning rate of 0.001, batch size of 32 and 50 epochs, monitoring validation loss to prevent overfitting through early stopping (patience=10). Hyperparameters for both CNN (e.g. filter sizes) and random forest (e.g. number of trees) are tuned using grid search with 5-fold cross-validation. Training occurs on a GPU-enabled system for CNN acceleration, while random forest training is central processing unit (CPU)-based for efficiency. This process ensures the model generalises well, similar to approaches in Shambharkar et al. (2023).^[8]

Evaluation metrics

Model performance is assessed using standard classification metrics: accuracy (overall correctness), sensitivity (true positive rate and crucial for early detection), specificity (true negative rate) and F1-score (harmonic mean of precision and recall, effective for imbalanced classes). These metrics provide a holistic view of the model’s diagnostic capabilities, aligning with evaluations in related studies (Suman et al., 2024).^[7]