The growing interest in hydrogen (H₂) has motivated process engineers and industrialists to investigate the potential of liquid hydrogen (LH₂) storage. LH₂ is an essential component in the H₂ supply chain. Many researchers have studied LH₂ storage from the perspective of tank structure, boil-off losses, insulation schemes, and storage conditions. A few review studies have also been published considering LH₂ storage; however, most are simply collections of previous articles. None of these review articles have critically evaluated the research articles. In this review study, recent reports, conceptual studies, and patents have been included and critically discussed. Further, challenges and recommendations have been listed based on the literature review. Our results suggest that the multi-layer insulation scheme and integrated refrigeration system can effectively reduce boil-off losses. However, boil-off losses from storage tanks during transportation are the least discussed and must be addressed. The cost of an LH₂ storage tank is high, but it can be reduced with advancements in materials and the utilization of latest technologies. The present challenges and future directions for LH₂ storage include minimizing and utilizing boil-off losses, improving insulation schemes, and ensuring cost-effective large-scale LH₂ storage. This review study can be fundamental for process engineers and new academic researchers to design energy-efficient and cost-effective LH₂ storage systems.

Polycystic ovary syndrome (PCOS) is a medical condition that impacts millions of women worldwide; however, due to a lack of public awareness, as well as the expensive testing involved in the identification of PCOS, 70% of cases go undiagnosed. Therefore, the primary objective of this study is to design an expert machine learning (ML) model for the early diagnosis of PCOS based on initial symptoms and health indicators; two datasets were amalgamated and preprocessed to accomplish this goal, resulting in a new symptomatic dataset with 12 attributes. An ensemble learning (EL) model, with seven base classifiers, and a deep learning (DL) model, as the meta-level classifier, are proposed. The hyperparameters of the EL model were optimized through the nature-inspired walrus optimization (WaO), cuckoo search optimization (CSO), and random search optimization (RSO) algorithms, leading to the WaOEL, CSOEL, and RSOEL models, respectively. The results obtained prove the supremacy of the designed WaOEL model over the other models, with a PCOS prediction accuracy of 92.8% and an area under the receiver operating characteristic curve (AUC) of 0.93; moreover, feature importance analysis, presented with random forest (RF) and Shapley additive values (SHAP) for positive PCOS predictions, highlights crucial clinical insights and the need for early intervention. Our findings suggest that patients with features related to obesity and high cholesterol are more likely to be diagnosed as PCOS positive. Most importantly, it is inferred from this study that early PCOS identification without expensive tests is possible with the proposed WaOEL, which helps clinicians and patients make better informed decisions, identify comorbidities, and reduce the harmful long-term effects of PCOS.

This narrative review presents a comprehensive and state-of-the-art synthesis of how machine learning (ML) is transforming public health through enhanced prediction, personalized treatment, real-time surveillance, and intelligent resource optimization. Drawing from 170 peer-reviewed studies published up to 2024/2025, this work uniquely integrates cross-domain insights spanning disease outbreak forecasting, genomic data analysis, personalized medicine, mental health monitoring, and public health infrastructure planning. The novelty of this review lies in its multidimensionality. It merges technical efficacy, ethical challenges, and future trends into a unified narrative. Our findings show substantial performance gains across domains: for example, ML models such as LightGBM, GRU neural networks, and LSTM achieved disease prediction accuracies ranging from 88 to 95%. In genomics, ML methods enabled nuanced disease subtype discovery and improved the accuracy of cancer risk assessment and pharmacogenomic modeling. Mental health prediction systems based on NLP and wearable data delivered up to 91% accuracy in stress and depression detection, while hospital resource forecasting models using deep learning minimized errors in predicting emergency admissions. Ethically, this review surfaces critical issues, including algorithmic bias, data privacy concerns in mental health analytics, and the interpretability of black-box models used in outbreak surveillance. A forward-looking discussion identifies future priorities such as the integration of multi-omics data, deployment of explainable AI, and equitable data inclusion frameworks. This review stands out by not only cataloguing applications but also offering a systems-level perspective on how ML can equitably and ethically scale to support public health strategies globally. It is among the first narrative reviews to concurrently evaluate ML’s predictive power, ethical constraints, and domain-specific improvements across all core pillars of public health.

Waste management handles all kinds of waste, including household, industrial, municipal, organic, biomedical, biological, and radioactive wastes. People still face challenges in proper disposal methods for different types of waste, including landfill-bound items, recyclable materials, and biodegradable waste. Inadequate waste management poses a significant and multifaceted global challenge. The conventional method of segregating waste is a time-consuming and ineffective method that wastes human power and money. To address this issue in real time, sophisticated and sustainable waste management systems need to be implemented. The latest advancements in computer vision and deep learning offer efficient solutions for effective recycling and waste management. Existing deep learning models exhibited various limitations, such as detection accuracy and computational inefficiency, particularly when dealing with objects of varying sizes and exhibiting high degrees of visual similarity. These limitations generate various challenges in effectively capturing and representing the nuanced features of visually similar objects. To address this problem, we proposed the stacking of an enhanced Swin Transformer, improved ConvNeXt, and a spatial attention mechanism. The enhanced Swin transformers incorporate two key components- hierarchical feature extraction and shifting window mechanism to extract the global features from the garbage images effectively. The shifting window mechanism extracts the most important features from various regions of the images to identify the objects. In contrast, the hierarchical feature extraction captures long-range dependencies within the image to effectively identify different types of garbage. The improved ConvNext block with optimized parameterization extracts the local features of the image. This enhanced feature extraction capability enables the model to effectively discern fine-grained details of individual garbage particles, such as shape, texture, and subtle variations in color and appearance, leading to more accurate classification results. When we evaluated the performance of the proposed model using the publicly available Garbage Classification dataset, it attained 98.97% accuracy, 98.42% Precision, and 98.61% Recall. Due to its lightweight and low computational time and power, the proposed model surpasses the existing state-of-the-art deep learning models.

In the context of lifestyle changes, stress and other environmental factors have resulted in the sudden hike in dementia globally. This necessitates investigations with respect to every horizon of the due cause for it; further on, the diagnosis and treatments can be advanced with the aid of technology. This work attempts to conduct one such investigation on dementia diagnosis based on EEG signals. The implementation is carried out under three different verticals. Firstly, a conventional machine learning model was developed post-pre-processing, and feature extraction from the power spectral density was done using a Random Forest classifier. Second, 1D Convolutional Neural Networks models are developed, and pre-processed EEG signals are fed as input. Third, stacked spectrogram images are computed from decomposed EEG signals and are fed to 2D CNN models for classification. The investigations are performed on three different benchmark datasets. Across three datasets, the class labels include cognitively normal, frontotemporal dementia, mild cognitive impairment, and Alzheimer’s. The study offers a comparative evaluation across three distinct datasets, illustrating that deep learning models, particularly 1D and 2D CNNs, consistently outperform conventional methods in recognizing subtle EEG signal patterns linked to neurodegenerative conditions. For instance, in Dataset 1, the 2D CNN achieved the highest accuracy of 91.13%, surpassing the Random Forest model’s 84.78% accuracy. Nevertheless, the investigation also points out challenges in Dataset 3, indicating the necessity for further model optimization tailored to specific datasets. Statistical tests validate the findings. This study showcases a comparative investigation of the potential of deep learning models vs. conventional classifiers in clinical environments for the early diagnosis of dementia.

Breast cancer ranks as the second most prevalent cancer globally and is the most frequently diagnosed cancer among women; therefore, early, automated, and precise detection is essential. Most AI-based techniques for breast cancer detection are complex and have high computational costs. Hence, to overcome this challenge, we have presented the innovative LightweightUNet hybrid deep learning (DL) classifier for the accurate classification of breast cancer. The proposed model boasts a low computational cost due to its smaller number of layers in its architecture, and its adaptive nature stems from its use of depth-wise separable convolution. We have employed a multimodal approach to validate the model’s performance, using 13,000 images from two distinct modalities: mammogram imaging (MGI) and ultrasound imaging (USI). We collected the multimodal imaging datasets from seven different sources, including the benchmark datasets DDSM, MIAS, INbreast, BrEaST, BUSI, Thammasat, and HMSS. Since the datasets are from various sources, we have resized them to the uniform size of 256 × 256 pixels and normalized them using the Box-Cox transformation technique. Since the USI dataset is smaller, we have applied the StyleGAN3 model to generate 10,000 synthetic ultrasound images. In this work, we have performed two separate experiments: the first on a real dataset without augmentation and the second on a real + GAN-augmented dataset using our proposed method. During the experiments, we used a 5-fold cross-validation method, and our proposed model obtained good results on the real dataset (87.16% precision, 86.87% recall, 86.84% F1-score, and 86.87% accuracy) without adding any extra data. Similarly, the second experiment provides better performance on the real + GAN-augmented dataset (96.36% precision, 96.35% recall, 96.35% F1-score, and 96.35% accuracy). This multimodal approach, which utilizes LightweightUNet, enhances the performance by 9.20% in precision, 9.48% in recall, 9.51% in F1-score, and a 9.48% increase in accuracy on the combined dataset. The LightweightUNet model we proposed works very well thanks to a creative network design, adding fake images to the data, and a multimodal training method. These results show that the model has a lot of potential for use in clinical settings.