Chronic disease prediction is a critical task in healthcare. Existing studies fulfil this requirement by employing machine learning techniques based on patient features, but they suffer from high dimensional data problems and a high level of bias. We propose a framework for predicting chronic disease based on Graph Neural Networks (GNNs) to address these issues. We begin by projecting a patient-disease bipartite graph to create a weighted patient network (WPN) that extracts the latent relationship among patients. We then use GNN-based techniques to build prediction models. These models use features extracted from WPN to create robust patient representations for chronic disease prediction. We compare the output of GNN-based models to machine learning methods by using cardiovascular disease and chronic pulmonary disease. The results show that our framework enhances the accuracy of chronic disease prediction. The model with attention mechanisms achieves an accuracy of 93.49% for cardiovascular disease prediction and 89.15% for chronic pulmonary disease prediction. Furthermore, the visualisation of the last hidden layers of GNN-based models shows the pattern for the two cohorts, demonstrating the discriminative strength of the framework. The proposed framework can help stakeholders improve health management systems for patients at risk of developing chronic diseases and conditions.
Graph machine-learning (ML) methods have recently attracted great attention and have made significant progress in graph applications. To date, most graph ML approaches have been evaluated on social networks, but they have not been comprehensively reviewed in the health informatics domain. Herein, a review of graph ML methods and their applications in the disease prediction domain based on electronic health data is presented in this study from two levels: node classification and link prediction. Commonly used graph ML approaches for these two levels are shallow embedding and graph neural networks (GNN). This study performs comprehensive research to identify articles that applied or proposed graph ML models on disease prediction using electronic health data. We considered journals and conferences from four digital library databases (i.e., PubMed, Scopus, ACM digital library, and IEEEXplore). Based on the identified articles, we review the present status of and trends in graph ML approaches for disease prediction using electronic health data. Even though GNN-based models have achieved outstanding results compared with the traditional ML methods in a wide range of disease prediction tasks, they still confront interpretability and dynamic graph challenges. Though the disease prediction field using ML techniques is still emerging, GNN-based models have the potential to be an excellent approach for disease prediction, which can be used in medical diagnosis, treatment, and the prognosis of diseases.
Artificial intelligence is changing the practice of healthcare. While it is essential to employ such solutions, making them transparent to medical experts is more critical. Most of the previous work presented disease prediction models, but did not explain them. Many healthcare stakeholders do not have a solid foundation in these models. Treating these models as ‘black box’ diminishes confidence in their predictions. The development of explainable artificial intelligence (XAI) methods has enabled us to change the models into a ‘white box’. XAI allows human users to comprehend the results from machine learning algorithms by making them easy to interpret. For instance, the expenditures of healthcare services associated with unplanned readmissions are enormous. This study proposed a stacking-based model to predict 30-day hospital readmission for diabetic patients. We employed Random Under-Sampling to solve the imbalanced class issue, then utilised SelectFromModel for feature selection and constructed a stacking model with base and meta learners. Compared with the different machine learning models, performance analysis showed that our model can better predict readmission than other existing models. This proposed model is also explainable and interpretable. Based on permutation feature importance, the strong predictors were the number of inpatients, the primary diagnosis, discharge to home with home service, and the number of emergencies. The local interpretable model-agnostic explanations method was also employed to demonstrate explainability at the individual level. The findings for the readmission of diabetic patients could be helpful in medical practice and provide valuable recommendations to stakeholders for minimising readmission and reducing public healthcare costs.
Purpose: Disease risk prediction poses a significant and growing challenge in the medical field. While researchers have increasingly utilised machine learning (ML) algorithms to tackle this issue, supervised ML methods remain dominant. However, there is a rising interest in unsupervised techniques, especially in situations where data labels might be missing — as seen with undiagnosed or rare diseases. This study delves into comparing unsupervised ML models for disease prediction. Methods: This study evaluated the efficacy of seven unsupervised algorithms on 15 datasets, including those of heart failure, diabetes, and breast cancer. It used six performance metrics for this comparison. They are Adjusted Rand Index, Adjusted Mutual Information, Homogeneity, Completeness, V-measure and Silhouette Coefficient. Results: Among the seven unsupervised ML methods, the DBSCAN (Density-based spatial clustering of applications with noise) showed the best performance most times (31), followed by the Bayesian Gaussian Mixture (18) and Divisive clustering (15). No single model consistently outshined others across every dataset and metric. The study emphasises the crucial role of model and performance measure selections based on application-specific needs. For example, DBSCAN excels in Homogeneity, Completeness and V-measure metrics. Conversely, the Bayesian Gaussian Mixture is good in the Adjusted R and Index metric. The codes used in this study can be found at https://github.com/haohuilu/unsupervisedml/. Conclusion: This research contributes deeper insights into the unsupervised ML applications in healthcare and encourages further investigations into model selection. Subsequent studies could harness genuine disease records for a more nuanced comparison and evaluation of models.
16 July 2025 in Artificial Intelligence in Medicine
Large Language Models (LLMs) are increasingly deployed in web-based medical advice applications, offering scalable and accessible healthcare solutions. However, their outputs often reflect demographic biases, raising concerns about fairness and equity for vulnerable populations. In this work, we propose FairMed, a framework designed to mitigate biases in LLM-generated medical advice through fine-tuning and prompt engineering strategies. We evaluate FairMed using language-based and content-level metrics across demographic groups on publicly available (MedQA), synthetic (Synthea), and private (CBHS) datasets. Experimental results demonstrate consistent improvements over Llama3 - Med42, as well as over the zero-shot prompting baseline. For instance, in sentiment analysis for gender groups using MedQA, FairMed with Descriptive Prompting reduces the Statistical Parity Difference (SPD) from 0.0902 to 0.0658, improves the Disparate Impact Ratio from 1.1916 to 1.1566, and decreases the Kullback-Leibler Divergence from 0.0045 to 0.0024. Similarly, in directive language evaluation for gender groups using Synthea, SPD improves from 0.1056 to nearly zero, achieving near-perfect parity. On the CBHS dataset, FairMed with Descriptive Prompting increases Diagnostic Recommendation Divergence (DRD) for race groups from 0.9530 to 0.9848, indicating improved group-specific tailoring, while reducing the Action Disparity Index (ADI) from 0.0857 to 0.0469 and Referral Frequency Parity (RFP) from 0.0791 to 0.0511, reflecting enhanced fairness. These findings highlight FairMed's effectiveness in addressing demographic disparities and promoting equitable healthcare guidance through web technologies. This framework contributes to building trustworthy and inclusive systems for delivering medical advice by ensuring fairness in sensitive applications.
The Delta variant of COVID-19 has been found to be extremely difficult to contain worldwide. The complex dynamics of human mobility and the variable intensity of local outbreaks make measuring the factors of COVID-19 transmission a challenge. The inter-suburb road connection details provide a reliable proxy of the moving options for people between suburbs for a given region. By using such data from Greater Sydney, Australia, this study explored the impact of suburban road networks on two COVID-19-related outcomes measures. The first measure is COVID-19 vulnerability, which gives a low score to a more vulnerable suburb. A suburb is more vulnerable if it has the first COVID-19 case earlier and vice versa. The second measure is COVID-19 severity, which is proportionate to the number of COVID-19-positive cases for a suburb. To analyze the suburban road network, we considered four centrality measures (degree, closeness, betweenness and eigenvector) and core–periphery structure. We found that the degree centrality measure of the suburban road network was a strong and statistically significant predictor for both COVID-19 vulnerability and severity. Closeness centrality and eigenvector centrality were also statistically significant predictors for COVID-19 vulnerability and severity, respectively. The findings of this study could provide practical insights to stakeholders and policymakers to develop timely strategies and policies to prevent and contain any highly infectious pandemics, including the Delta variant of COVID-19.