Artificial Intelligence Applications in Public Health: 2nd Edition

Artificial intelligence (AI) is assuming an increasingly important role in public health, where the scale, heterogeneity, and temporal dynamics of health-related data often exceed the capacity of conventional analytic approaches. Recent scholarship has shown that AI can strengthen epidemic intelligence, support earlier detection of health threats, improve predictive modeling, and enhance evidence generation for public health decision-making. For example, AI-based systems have been recognized as useful for epidemic monitoring and alerting, particularly when rapid integration of multiple data streams is required [1]. At the same time, broader reviews have emphasized their growing relevance for health policy, service planning, and data-driven governance [2]. The literature makes clear that the expansion of AI in health-related domains must be accompanied by careful attention to interpretability, bias, privacy, accountability, and implementation constraints, especially when computational outputs inform population-level interventions or resource allocation [3]. These considerations underscore that the significance of AI in public health lies not only in technical performance but also in its capacity to produce actionable, trustworthy, and context-sensitive insights for complex health systems.

Against this background, the Special Issue “Artificial Intelligence Applications in Public Health: 2nd Edition” is both timely and well justified. This volume is explicitly presented as the second edition of the collection, continuing and extending the thematic focus of the earlier Special Issue [4]. Its stated scope highlights the transformative potential of AI for disease surveillance, outbreak prediction, health systems optimization, and personalized health interventions, while also emphasizing the broader contribution of computation to public health decision-making. The continuation of the topic into a second volume reflects the sustained development of this interdisciplinary field and the growing range of research questions that now fall within the domain of AI-enabled public health. Rather than treating AI as a narrow set of clinical tools, this Special Issue situates it within a broader methodological landscape that includes epidemiological modeling, medical data analysis, decision support, and computational approaches to public health management.

The papers published in this second volume illustrate the breadth of contemporary research at the intersection of AI and public health. They address infectious disease forecasting, epidemic dynamics, clinical decision support, mental health analytics, biomedical signal analysis, data quality, project management, and emerging computational paradigms for medical applications. Several contributions focus on infectious disease modeling, including the comparative evaluation of neural network architectures for forecasting, smoothing methods to improve COVID-19 time series prediction, and the analysis of epidemic behavior under the influence of hidden infections, re-infections, and demographic processes. Other studies examine clinically oriented predictive tasks, including the diagnosis of severe COVID-19, the impact of missing data imputation on supervised learning performance, and survival prediction in allogeneic bone marrow transplantation under small data conditions. The Special Issue also extends beyond infectious disease and acute clinical prediction to include multimodal depression detection, scalable clustering of complex electrocardiographic data, AI-supported management of public health projects in unstable environments, and exploratory hybrid quantum–classical neural network models for medical use. These contributions demonstrate that AI applications in public health are methodologically diverse and substantively wide-ranging, spanning both population-level and patient-oriented problems while highlighting the increasing maturity of computational approaches in this field.

The paper “Multimodal Data Fusion for Depression Detection Approach” (Contribution 1) addresses the task of automatic depression detection from multimodal interview data by integrating textual and audio information. The authors propose and compare two multimodal deep learning architectures, early fusion and late fusion, built on CNN, Bi-LSTM, and self-attention components, and evaluate them on the DAIC-WOZ dataset supplemented with additional depression cases from EDAIC-WOZ to mitigate class imbalance. Their preprocessing pipeline includes transcript correction, removal of interviewer utterances, augmentation of minority-class samples, extraction of audio features such as Mel spectrograms, MFCCs, and spectral contrast, and TF–IDF-based textual representation. The results show that the early fusion model achieved superior performance, reaching 0.93 accuracy on the validation set and 0.85 accuracy with an F1-score of 0.79 for the depression class on the test set, thereby outperforming the late fusion alternative. The authors conclude that the earlier integration of modalities allows the model to capture cross-modal depressive cues more effectively and that such multimodal systems have practical potential for early depression screening, while still requiring careful clinical interpretation and appropriate ethical safeguards.

The paper “COVID-19 Data Analysis: The Impact of Missing Data Imputation on Supervised Learning Model Performance” (Contribution 2) investigates the impact of missing data imputation on the performance of supervised machine learning models for COVID-19 case classification using a hospital dataset from the Concepción Department of Paraguay. The authors compare five classifiers (Random Forest, Artificial Neural Network, Support Vector Machine, Logistic Regression, and Decision Tree) combined with four imputation techniques (PMM via MICE, Random Forest, KNN, and XGBoost-based imputation) across multiple levels of missingness, evaluating performance using accuracy, AUC, F1-score, MCC, sensitivity, and specificity. The results show that Random Forest was the most robust model overall, maintaining strong performance even at high levels of missingness, with baseline values of 0.826 accuracy and 0.902 AUC, and only limited degradation after imputation. In contrast, SVM exhibited a marked specificity–sensitivity trade-off, whereas ANN proved more sensitive to changes in imputation-related data quality. The authors conclude that the choice of imputation method substantially affects predictive performance and that ensemble-based approaches, particularly Random Forest, offer the most reliable solution for incomplete medical datasets, especially in resource-constrained healthcare settings.

The paper “Cascade-Based Input-Doubling Classifier for Predicting Survival in Allogeneic Bone Marrow Transplants: Small Data Case” (Contribution 3) addresses the task of predicting survival after allogeneic bone marrow transplantation under small data and class-imbalance conditions, where conventional machine learning models often lose reliability. To improve classification performance, the authors propose a cascade-based input-doubling classifier in which a Probabilistic Neural Network (PNN) first generates intermediate predictions that expand the feature space, followed by a second machine learning model that performs the final classification on the augmented data. Four algorithmic variants were implemented, combining PNN with Random Forest, XGBoost, HistGradientBoosting, or a second PNN, with parameters optimized using the Dual Annealing method and performance assessed by five-fold cross-validation. The results show that the PNN + Random Forest configuration achieved the best performance, reaching 0.97 accuracy, 0.98 F1-score, 0.95 MCC, and 0.95 Cohen’s kappa, outperforming both the baseline input-doubling method and a range of standard classifiers, including classical PNN and Random Forest. The authors conclude that the proposed cascade-based enhancement substantially improves predictive accuracy for survival estimation in transplantation medicine and has practical value for clinical decision support in settings with limited medical data, despite a higher computational cost than simpler baseline models.

The paper “How Re-Infections and Newborns Can Impact Visible and Hidden Epidemic Dynamics?” (Contribution 4) addresses the task of modeling visible and hidden epidemic dynamics under the joint influence of re-infections, newborns, and unregistered cases, aiming to improve long-term epidemic prediction beyond the assumptions of the classical SIR framework. The author proposes a generalized system of five differential equations that separates infectious and removed populations into visible and hidden compartments, incorporates waning immunity and new susceptible inflow due to births, and examines equilibrium points, quasi-equilibria, and numerical solutions in relation to observed COVID-19 dynamics in South Korea and Austria and the pertussis epidemic in England. The results show that re-infections, hidden infections, and newborn inflow can prevent convergence to epidemic extinction, generate stable quasi-equilibrium states, and produce repeated epidemic waves. Notably, the pertussis simulations suggest that newborns alone can sustain recurrent outbreaks, with the next peak in England projected for 2031. The author concludes that epidemics may become effectively endless when hidden transmission and demographic renewal are taken into account, and that the proposed model offers a useful framework for estimating both visible and hidden epidemic burdens, provided that robust parameter-identification procedures are applied.

The paper “Smoothing Techniques for Improving COVID-19 Time Series Forecasting Across Countries” (Contribution 5) addresses the task of improving COVID-19 time series forecasting across countries by examining how different smoothing techniques affect predictive accuracy under short- and medium-term horizons. Using weekly case data from Ukraine, Bulgaria, Slovenia, and Greece, the authors compare four preprocessing methods (rolling mean, exponentially weighted moving average, Kalman filter, and STL decomposition) in combination with four forecasting models: LSTM, Temporal Fusion Transformer, XGBoost, and LightGBM. The experimental design evaluates all model–smoother combinations using RMSE, MAE, and MAPE. Then, it applies a two-way ANOVA to assess the statistical contributions of model architecture and smoothing choice. The results show that smoothing improves forecast stability, particularly for neural models, with LSTM and TFT paired with STL or rolling mean performing best for 3-month forecasts. In contrast, XGBoost shows greater robustness at 6-month horizons across selected countries. The authors conclude that model choice is the principal determinant of forecast accuracy, while smoothing remains a useful complementary step for stabilizing epidemic signals and tailoring forecasting pipelines to specific data-quality conditions and prediction horizons.

The paper “Scalable Clustering of Complex ECG Health Data: Big Data Clustering Analysis with UMAP and HDBSCAN” (Contribution 6) investigates the task of unsupervised identification of latent cardiac risk profiles from large-scale ECG and heart rate variability data, to support early heart failure risk stratification in both clinically healthy individuals and hospitalized cardiac patients. The authors analyze two high-dimensional datasets containing 153 ECG-derived features each and apply a combined UMAP + HDBSCAN framework, testing both Euclidean and Manhattan distance metrics and using a cross-dataset clustering strategy in which models trained on one population are projected onto the other to assess structural divergence and latent predisposition patterns. The results reveal stable multi-cluster structures in both cohorts, with certain clusters in the patient data exhibiting more pronounced risk-related signatures, and with cross-dataset projection showing that some “healthy” individuals align with cluster profiles more typical of cardiac pathology. Among the most discriminative features were the QRS angle in the frontal plane, DFA, HF power, ST dislocation, and normalized T-wave amplitude, which differentiated lower- and higher-risk groups across clustering settings. The authors conclude that scalable unsupervised clustering of ECG-derived features is feasible and may offer a promising, non-invasive approach to early cardiac risk stratification and personalized monitoring. However, the clinical significance of these clusters requires further validation in dedicated medical studies.

The paper “Successful Management of Public Health Projects Driven by AI in a BANI Environment” (Contribution 7) addresses the challenge of managing public health projects in a BANI environment through an AI-driven framework that responds to the combined pressures of epidemic spread, misinformation, panic, and war-related disruption. To this end, the authors develop a coupled SEIR–Infodemic–Panicdemic model adapted to the emergency context of Kharkiv, Ukraine, incorporating time-varying factors such as population displacement, healthcare disruption, and periodic conflict events, and use it to conduct a sensitivity analysis alongside qualitative and quantitative risk–opportunity assessments over a 1500-day simulation horizon. The results show that disease transmission, displacement, misinformation, and infrastructure damage substantially intensify public health crises, with some risk scenarios increasing infectious peaks by up to 28.3%, while AI-enabled opportunities such as fact-checking, panic reduction, and infrastructure recovery reduce misinformation, panic, and infection peaks by clinically meaningful margins. The authors conclude that AI can play a transformative role in crisis public health management by supporting predictive analytics, real-time monitoring, decision support, and adaptive intervention planning, thereby improving resilience and sustainability in highly volatile and uncertain environments.

The paper “Management of Severe COVID-19 Diagnosis Using Machine Learning” (Contribution 8) addresses the task of predicting COVID-19 severity using supervised machine learning on an integrated set of demographic, clinical, biochemical, and genetic variables, aiming to identify the most informative prognostic markers for severe disease. The authors analyze a dataset of 226 patients with confirmed COVID-19 and 68 candidate features, perform preprocessing with missing data handling and removal of highly correlated variables, and then compare multiple classifiers, including Logistic Regression, k-Nearest Neighbors, Decision Trees, Random Forest, Gradient Boosting, Bagging, Naïve Bayes, Support Vector Machines, and Ensemble Tree methods, under stratified five-fold cross-validation, using accuracy and AUC-ROC as evaluation metrics alongside feature-importance analysis. The results show that IL-6, depression/pneumonia, LDL cholesterol, AST, platelet count, lymphocyte count, and ALT were the strongest predictors of severity. At the same time, ensemble-based classifiers, particularly ExtraTreesClassifier, HistGradientBoostingClassifier, BaggingClassifier, and GradientBoostingClassifier, achieved the highest predictive performance, with the optimized dataset slightly improving stability and accuracy. The authors conclude that machine learning, especially ensemble methods, provides a highly effective framework for diagnosing severe COVID-19. At the same time, interpretable models such as decision trees remain valuable in clinical settings because they combine strong predictive capacity with transparent, rule-based reasoning.

The paper “Impact of Scaling Classic Component on Performance of Hybrid Multi-Backbone Quantum–Classic Neural Networks for Medical Applications” (Contribution 9) examines how the scale of the classical backbone affects the performance of hybrid quantum–classical neural networks for image classification in medical applications. The authors construct hybrid models by integrating a four-qubit parameterized quantum circuit with trainable Ry rotations into several scaled versions of LCNet and evaluate them against their classical counterparts on both CIFAR-10 and DermaMNIST, using stratified 5-fold cross-validation, out-of-fold predictions, and metrics including accuracy, macro- and micro-AUC, per-class accuracy, and fold-to-fold stability. The results show that hybrid LCNet models provide their most consistent benefits at small and medium scales, where they improve both predictive accuracy and robustness; on DermaMNIST, for example, compact hybrid models achieved gains of up to 6–7 percentage points in accuracy and more than 3 points in AUC, while also reducing performance variability across folds. However, as backbone size increased, the classical architectures regained superiority, indicating that the advantages of quantum integration diminish as classical capacity increases. The authors conclude that the effectiveness of hybrid quantum–classical models depends fundamentally on architectural compatibility and model scale. Lightweight backbones such as LCNet offer the most promising foundation for practical, resource-efficient hybrid systems in data-limited and clinically relevant medical settings.

The paper “Improving the Accuracy of Infectious Disease Forecasts Based on Comparing Neural Network Architectures” (Contribution 10) addresses the task of improving infectious disease forecasting accuracy by comparing the performance of several neural network architectures on national surveillance time series from Ukraine. The authors compiled and normalized monthly data on 57 infectious diseases for the period December 2016 to January 2024, quantified war-related shifts in morbidity, and then evaluated RNN, Block RNN, N-BEATS, TCN, Transformer, and N-HiTS under two training strategies—joint training on all disease-specific time series and separate training for each disease—using the previous 24 months to predict the next 6 months and assessing performance with MAPE, MAE, and RMSE. The results show that N-BEATS achieved the best overall forecasting accuracy, with an average MAPE of 63.56% in the multi-series setting, and that models trained jointly on multiple disease series consistently outperformed those trained on single series, suggesting the presence of informative shared temporal structure across infections. The authors conclude that neural forecasting models, particularly N-BEATS, can support public health decision-making in emergency settings. However, their performance remains limited for diseases without clear seasonal or trend patterns and would likely benefit from the inclusion of exogenous variables and more diverse datasets in the future.