The COVID-19 pandemic led to the acceleration of the deep learning (DL) usage. This study focuses on the use of binary classification and multi-class classification in COVID-19 detection and pneumonia detection. Even though the models demonstrated high accuracy, the inspection shows issues related to the dataset bias, limitations and the robustness of the validation techniques. There is a progressive shift in the trajectory toward more robust methods with incorporation of Explainable AI (XAI). This review sheds light on methods, performance and limitations of previous models.

White blood cells (WBCs) are integral components of the immune system, playing a pivotal role in defending the body against various diseases and infections. Accurate quantitative and qualitative analysis of different WBC types is essential for the effective diagnosis and treatment of numerous medical conditions. Counting and classifying these cells are fundamental steps in blood sample examination and testing. This study focuses on evaluating three convolutional neural network (CNN)-based architectures – ResNet-50, VGG-16, and a classic CNN – for classifying blood cell image categories using a curated dataset from Kaggle. We trained and evaluated each model’s performance using accuracy curves, confusion matrices, and classification reports. Among the tested architectures, the ResNet-50 model achieved the highest validation accuracy of approximately 80%, followed by VGG-16 at 79.6% and the classic CNN at 76.2%. Both VGG-16 and the classic CNN exhibited significant overfitting, with large gaps between training and validation accuracy. These findings highlight the challenges of image classification on imbalanced datasets and suggest directions for future improvements through data augmentation and architectural refinements.

Parkinson's disease (PD) lacks objective, low-cost early diagnostic tools in routine clinical practice. Handwriting analysis offers a promising non-invasive biomarker, as PD-related motor impairments measurably degrade pen-trace quality before overt symptoms appear. This paper presents a systematic 40-run factorial benchmark evaluating five preprocessing pipelines raw RGB (P1), grayscale (P2), CLAHE (P3), CLAHE + Gaussian (P4), and Canny edge detection (P5) combined with four CNN architectures (MobileNetV2, ResNet18, EfficientNetB0, VGG16) across spiral and wave modalities from the Parkinson's Augmented Handwriting Dataset (1,020 images). Results show that minimal preprocessing consistently yields the best performance: MobileNetV2 + P1 achieves 90.67% accuracy and $\mathbf{A U C}=\mathbf{0. 9 7 0}$ on spiral drawings, while EfficientNetB0 + P1 and VGG16 + P3 both attain 91.33% accuracy with AUC = 0.9785 and 0.9765 on wave drawings. Canny edge detection uniformly degrades all architectures, confirming that ImageNet-pretrained CNNs perform optimally on minimally processed inputs.

The advancements that have been made in Artificial Intelligence (AI) make it one of the fastest growing areas of research in science and technology. AI’s promise for Medicine (especially Oncology) is its ability to increase the accuracy of diagnosis, improve treatment strategies, assist in clinical decision-making, and minimize human error. As a result of advances in Artificial Intelligence (AI), there is a major opportunity to provide more accurate diagnoses and greater personalization for managing cancer, using innovative AI-driven solutions that have shown promise in supporting early detection and prognosis prediction. In this manuscript, we summarize recent research findings on how machine learning (ML) techniques are useful in diagnosing lung cancer, and how they may improve prognosis prediction and clinical decision support for lung cancer patients and their families. Based on this information, we developed a Convolutional Neural Network model to classify lung cancer images that were made publicly available as the IQ-OTH/NCCD Lung Cancer Dataset, consisting of 1190 labeled CT scans classified as either normal, benign, or malignant. The dataset consists of multiple grayscale CT scans collected from a number of different medical centers and manually annotated to form a highly realistic and heterogeneous dataset suitable for training supervised learning algorithms. Our CNN-based technique is able to capture spatial and textural features from CT scans to accurately classify lung nodules. The purpose of this interface is to allow end-users to bridge the gap from complex algorithms to clinical workflows easily and to introduce AI into oncology than previously possible.

Parkinson’s disease (PD) manifests through both motor and non-motor symptoms (NMS), the latter playing an increasingly recognized role in early and differential diagnosis. Although recent works using the Parkinson’s Disease Smartwatch (PADS) dataset have focused primarily on multimodal fusion strategies combining wearable sensor data with clinical questionnaire responses, the standalone diagnostic capability of NMS questionnaires has not been thoroughly investigated. This paper presents a systematic evaluation of PD classification relying exclusively on the $\mathbf{3 0}$-item PDNMS questionnaire from the PADS dataset, without any smartwatch sensor data. We employ three supervised learning classifiers: CatBoost, Random Forest, and XGBoost combined with SelectKBest feature selection (ANOVA F-test, $\mathrm{k}=17$), SMOTE based class balancing on training data only, and 5-fold stratified cross-validation with hyperparameter grid search. The questionnaire only approach achieves balanced accuracies of 93.24% for PD vs. Healthy Controls (HC) and 78.22% for PD vs. Differential Diagnosis (DD), equaling or surpassing prior questionnaire only baselines and rivaling sensor-based multimodal approaches. These results establish NMS questionnaires as a powerful, cost-effective, and hardware-free alternative for large scale PD screening.

Parkinson’s disease (PD) is a progressive neurodegenerative disorder for which early and objective diagnosis remains challenging, primarily due to the subjective nature of conventional clinical assessment. Recent advances in wearable sensor technology including smartwatches and inertial measurement units (IMUs) now enable continuous and unobtrusive monitoring of motor symptoms, opening new avenues for objective, data-driven diagnostic systems. In this paper, we provide a structured overview of machine learning (ML) and deep learning (DL) algorithms applied to wearable sensor data for PD diagnosis and motor symptom evaluation. We systematically examine classical ML approaches, deep learning architectures, and hybrid methods, analyzing their performance across three clinical tasks: PD detection, symptom severity estimation, and differential diagnosis from other movement disorders. Based on a selection of eleven peer-reviewed studies, our analysis highlights the complementary strengths of ML and DL, identifies key challenges such as intersubject variability, class imbalance, and lack of standardized evaluation protocols, and discusses promising directions for future research.

Artificial intelligence (AI) and machine learning algorithms are transforming quality audit processes in hospital pharmacies. While these technologies offer unprecedented capabilities for improving medication safety, detecting anomalies, and ensuring regulatory compliance, they also introduce new cybersecurity risks and pose ethical challenges. This article reviews the current landscape of AI-based audit methodologies, examines cybersecurity vulnerabilities specific to pharmaceutical environments, and synthesizes evidence-based recommendations for implementing secure and reliable systems. We conducted a systematic analysis of peer-reviewed literature and institutional reports published between 2023 and 2025, identifying key themes such as anomaly detection, predictive analytics, data privacy concerns, and regulatory compliance challenges. Our findings highlight the critical importance of interdisciplinary approaches combining technical safeguards, regulatory frameworks, and organizational governance to harness the potential of AI in pharmaceutical auditing while mitigating inherent risks.

The ability to detect cancer at an early stage relies more than ever on developing ultrasensitive analytical tools that can identify low-abundance molecular markers present in the complex environments of our bodies. Surface-enhanced Raman spectroscopy (SERS) is widely accepted as one of the best techniques for label-free molecular fingerprinting; however, signal variability, background fluctuations, and lack of versatile computational workflows for automated classification of cancer biomarkers interfere with the utility of this technique. This manuscript describes the introduction of an integrated SERS-analyzers with intelligent algorithms for improving the detection of cancer biomarkers from SERS spectral data. We utilized two experimental datasets consisting of filtered and unfiltered samples and applied a standardized workflow comprised of noise–and baseline-correction and vector normalization. We then analyzed the results to compare the difference in spectral features and peak displacements caused by the filtration process. To identify the key discriminatory vibrational modes and reduce spectral redundancy, we employed principal component analysis (PCA) and anomaly detection techniques. The 30-kDa filtrates produced a unique set of spectral signatures that showed apparent changes in the prominent Raman peaks near 1588, 1605, and 2111, which are believed to correspond to specific molecular components common to cancer-associated biomarker pathways. Our study has demonstrated that the combination of selective filtration and intelligent processing methods significantly enhances the sensitivity of SERS for detecting cancer-related bio-chemical changes. The proposed framework will provide an avenue toward creating an intelligent network of sensors capable of providing real-time, noninvasive monitoring of cancer biomarkers.

Continuous electrocardiogram (ECG) monitoring requires practical engineering solutions that balance predictive performance, response time, and resource consumption in distributed wearable-oriented systems. We present a wearable-fog-cloud framework for continuous ECG decision orchestration in which wearable devices perform first-pass triage and the fog gateway evaluates patient-specific temporal criticality, uncertainty and resource admissibility, reserving cloud consultation for ambiguous or critical cases. The method explicitly distinguishes logical cloud demand from actual cloud execution and incorporates exact record-level decision memoization for repeated beat-level tuples derived from the same ECG record. The framework was evaluated in iFogSim as a controlled system-level simulation using synchronized wearable-side and cloud-side prediction streams derived from PTB-XL. We compared the proposed method with three baselines: wearable-only, cloud-only, and no-reuse. We also assessed scalability, parameter sensitivity, cache realism, component-level ablations, and bootstrap uncertainty at the record and patient levels. At the selected operating point, the method achieved a Matthews correlation coefficient of 0.6073, an F1-score of 0.8255, and an area under the receiver operating characteristic curve of 0.8803, with an average latency of 27.9 ms and an actual cloud offload ratio of 6.13%. Among logically cloud-resolved tuples, 84.66% were served through exact record-level cache reuse. The proposed framework improved predictive performance compared with wearable-only inference while maintaining low latency; compared with the no-reuse configuration, it achieved higher predictive performance with slightly lower real cloud execution. These results indicate that selective cloud escalation and exact record-level decision reuse can improve the trade-off between predictive quality and resource usage in a controlled ECG decision-orchestration framework, while prospective validation on native wearable or ambulatory ECG data remains necessary.

Voice analysis is a non-invasive method for Parkinson’s disease (PD) screening: vocal impairments are reported in $\mathbf{7 0}-\mathbf{9 0 \%}$ of patients. Machine learning methods applied to this task lack temporal explainability, which limits their use in clinical practice. This paper proposes a temporal attention mechanism for PD detection from sustained vowel recordings. Two features, micro-pause ratio and energy decay, are introduced to measure vocal instability not captured by standard acoustic measures; both reach statistical significance ($p=1.95 \times 10^{-21}$). At 500 ms resolution, the attention module shows that healthy speakers weight phonation onset ($0-0.5 \mathrm{~s}$), while PD speakers produce a bimodal weight profile with a second peak at $\mathbf{3. 5 - 5. 0} \mathrm{s}$, which corresponds to motor deterioration under sustained load. Attention weights are overlaid on the waveform so that clinicians can verify model decisions. Attention peaks beyond 3 s were found in 11.4% of PD patients and in no healthy controls, indicating a pattern that emerges only under sustained phonatory demand. On 831 recordings the model yields 80.84% accuracy, 86% sensitivity, and $\mathrm{AUC}=0.8441$.

Parkinson’s disease (PD) is routinely confirmed only after substantial neuronal loss, by which point motor deterioration is already underway. Voice, however, shows measurable signs of the disease much earlier speech impairments affect up to 90 % of patients and frequently precede frank motor decline by several years. Building on this observation, we designed ParkiVoiceNet, a deep learning system with two parallel analysis paths: one processes the raw waveform through stacked Conv1D layers followed by a BiLSTM and a self-attention mechanism; the other applies four Conv2D layers to the log Mel-spectrogram. The two paths are combined by late fusion before the output layer. We evaluated the system on 831 recordings from 65 speakers (Italian dataset) under a strict subject-wise five-fold protocol, obtaining 93.74 % accuracy, 97.81 % AUC, 94.04 % sensitivity, and 93.39% specificity. At 192130 parameters, it runs on an ordinary smartphone. To make the predictions usable in practice, we added three interpretability tools Grad-CAM heatmaps, tSNE projections, and attention-weight inspection that allow a clinician to see which acoustic features drove each decision, rather than relying on a bare confidence score.

This study presents a comprehensive framework for automatic speech recognition (ASR) tailored to dysarthric speakers, leveraging the newly developed UASPEECH Preprocessed Dataset. The original UASPEECH recordings were enhanced through systematic signal processing techniques, including FFT-based denoising, Hanning windowing, single-sided amplitude doubling, and amplitude normalization, resulting in clean, standardized 16-bit WAV audio at 44.1 kHz. Two modeling pipelines were explored: one using handcrafted acoustic features and another based on mel spectrogram representations. For the acoustic features pipeline, several classical and deep learning models were evaluated, including Multi-Layer Perceptrons (MLP_1–3), LSTM, GRU, BiLSTM, BiGRU, CNN_1D, and an attention-only model. Among these, BiLSTM achieved the best performance, with an accuracy and F1-score of 97.25

Artificial Intelligence forms the foundation of smart-health and continuous monitoring systems, offering automated diagnostics and patient triage. The chest X-ray remains the most commonly acquired imaging modality across the world, however the wide range of presentation, anatomical variation and the uneven access to radiologists makes thoracic interpretation a challenge. In this study we establish the first baseline for multi-pathology detection, utilizing the public Chest Radiograph at Diverse Institutes (CRADI), available on Zenodo and fully processed on the Kaggle platform for reproducibility. Instead of training a deep-learning model, we probe the discriminability of the dataset, putting an untrained ResNet-18 model through its paces. We show that, despite no learned representation, the baseline model yields a micro-AUC of 0.683 and non-random behaviour to several clinically relevant high-contrast thoracic conditions. Probing for pathology-specific properties reveals the variability of visual separability across the spectrum of 25 abnormalities and high-contrast phenotypes, revealing the inherent difficulty of multi-label thoracic detection and a need for approach for feature extraction. To complement the diagnostic baseline we present a simulated accuracy latency analysis in the overhead cross-section of lightweight CNNs vs. EfficientNet-like architectures and transformer based models, highlighting the latency-useful for smart-health and edge-AI deployment. We provide a benchmark-crude baseline for the CRADI and useful suggestions in designing lightweight AI networks in intelligent sensor networks.

Mortality associated with pulmonary tuberculosis (PTB) can be significantly reduced by early diagnosis. Chest X-rays are an effective means of detecting PTB, but accurate interpretation of these images requires the presence of an experienced radiologist, who are often in short supply in developing countries. Artificial intelligence, particularly convolutional neural networks (CNNs) in deep learning, offers a promising solution for early diagnosis of PTB. Exploring the potential of deep learning in image analysis, we investigated CNN algorithms to build a model capable of binary classification of these images, which can help predict the presence or absence of TB in chest radiographs. Using the concept of ensemble learning, we proposed a model of different CNN architectures optimized by deep learning. To train and validate the model, we used a dataset of chest radiographs. The results obtained in terms of classification are significant and encouraging.

Developing dependable monitoring systems is essential for ensuring the security and safety of patients. In the last decades, recent progress in IoT and embedded devices has enabled the conception of cost-effective real-time systems. Patient motion, which occurs owing to breath, cough, sneezing, or other activities, during the Computed Tomography (CT) scan is a serious authoritative complication that produces image degradation, elevation of radiation exposure, and repetition of scans. The study developed an automatic motion detection system by using sensors to develop such systems to ensure both patient safety and diagnostic accuracy. A prototype was created to establish real-time monitoring of patient mobility during computed tomography scans using Raspberry Pi. It uses advanced image processing techniques to detect motion and stop the emission of X-rays from the machine as soon as the motion of the patient is detected. The prototype has been assessed under several conditions, and values such as motion detection accuracy and response time were evaluated. The deployment of the prototype of the motion detection reduced kinetic blur, which is one of the main causes of repeated examinations. The system has been 90% accurate while detecting movements of patients significantly improving the quality of images produced from diagnostic facilities and less unnecessary exposure to radiation. The incorporation of real-time motion detection in CT scanners achieves greater image quality and reduction of patient radiation risks. The authors argue that such technologies should be adopted for clinical practice: in essence, they lead to superior patient care at reduced costs.

Respiratory diseases represent a major cause of morbidity and mortality worldwide, highlighting the need for rapid and accurate diagnostic tools. This article introduces an innovative method for the automatic detection of wheezes based on a convolutional neural network (CNN) model. By leveraging a database of respiratory sound recordings and advanced feature extraction techniques such as Mel-frequency cepstral coefficients (MFCC), our system achieves a 93

Parkinson's disease (PD) is a progressive neurodegenerative disorder that primarily impairs motor functions, leading to symptoms such as tremors and micrographia. Even though early identification of PD is crucial for effective intervention, existing methods of diagnosis are highly invasive and not very sensitive to the early stages of the diseases. The goal of this research is to determine if handwriting could be a non-invasive way to diagnose PD at an early stage. We employed a dataset of 3,264 hand-drawn waves and spirals to evaluate the performance of hybrid machine learning and deep learning models which included Support Vector Machine (SVM), Random Forest (RF), Visual Geometry Group-16 (VGG-16) and MobileNetV2. Combing SVM with VGG-16 for the task reached a stunning 99.00% accuracy for identifying PD, performing the best out of all tested models, demonstrating superior performance in the early identification of PD. The proposed approach not only outperforms existing diagnostic methods but also underscores the transformative potential of handwriting analysis tools in PD diagnosis, aiding in automatic PD detection and enhancing patient outcomes.

Parkinson's disease (PD) is a progressive neurological disorder that affects millions worldwide, leading to motor dysfunction and significant reductions in quality of life. Early diagnosis is pivotal for initiating timely treatment and improving long-term patient outcomes, yet existing diagnostic methods, which often rely on clinical evaluations and imaging, are prone to delays and varying accuracy. This study presents an innovative, non-invasive approach to early PD detection through the analysis of handwriting patterns, offering a potential alternative to traditional diagnostic techniques. Leveraging a publicly available and meticulously normalized handwriting dataset, our approach applies advanced data processing methods to identify subtle neuromotor impairments associated with PD. Through the integration of robust feature selection processes and cutting-edge machine learning models, we achieved a high accuracy rate of 83.02%, highlighting the method’s reliability. The findings suggest that this approach could significantly enhance early PD detection, leading to more personalized therapeutic strategies that align with the stages of disease progression and potentially delaying the onset of severe symptoms.

Blood sampling is a routine procedure in medical diagnostics, yet precise vein visualization methods remain limited. This project introduces a system designed to improve vein detection during blood collection. It relies on Near-Infrared (NIR) light, which interacts with the skin and highlights veins by taking advantage of hemoglobin’s infrared absorption properties. Using a Raspberry Pi and an infrared camera, image acquisition and processing are handled through MATLAB and Python algorithms, which allow real-time visualization of veins. The system has been tested on a database of infrared images of hands and arms, effectively enhancing vein contrast in real time. The display is connected to the Raspberry Pi, giving medical staff a visual guide. This technology aims to streamline procedures for healthcare professionals, including doctors, nurses, and medical students, particularly in high-volume settings like labs and blood transfusion centers where vein visualization is critical to patient care.

Parkinson’s disease (PD) is a neurodegenerative disorder (ND) with vocal impairments that complicate its differentiation from other neurological disorders. This study introduces a novel database of voice recordings from 40 PD patients and 20 patients with other neurological disorders, collected in both clinical and natural environments. A comprehensive set of 218 acoustic features was extracted, and machine learning algorithms—including Random Forest, Support Vector Machine, k-Nearest Neighbors, AdaBoost and others—were evaluated under three scenarios: no dimensionality reduction, linear Principal Component Analysis (PCA), and Non-linear Principal Component Analysis (NPCA). Using Leave-One-Subject-Out (LOSO) cross-validation, NPCA significantly enhanced classification performance, with algorithms achieving up to 95