To determine whether ATF3 polymorphisms can serve as genetic biomarkers for machine learning–based precision analgesia by establishing a genotype–phenotype association suitable for predictive modeling of postoperative opioid requirements. In a prospective cohort of 167 adults undergoing abdominal surgery, ATF3 SNPs rs3122721 and rs3125293 were genotyped. A structured dataset architecture was developed to represent genetic profiles as input features for supervised learning models, enabling translational analysis of genotype‑dependent opioid consumption over 72 h. Patients with homozygous genotypes of the ATF3 SNPs had significantly higher opioid requirements than non‑carriers, despite reporting similar subjective pain scores. This consistent genotype‑dependent pattern provided a clinically relevant phenotype suitable for integration into predictive algorithms. ATF3 genotyping offers a promising biomarker for computationally informed precision analgesia. By linking genomic variability to clinically meaningful outcomes within a structured clinical and genomic framework, this approach supports the future development of risk-stratified clinical decision-support systems to optimize postoperative pain management. Trial registration ChiCTR1900021991, registered 30 April 2019.
Medical datasets often lack high-quality annotated images, making it difficult to apply traditional deep learning techniques to create an unbiased model with strong generalization capabilities. To tackle this issue, we redefine the problem as a few-shot learning challenge and propose a meta-learning-based solution called “DerMeL.” This research presents a lightweight meta-learning algorithm designed to classify rare skin diseases in long-tailed distributions and resource-constrained environments. DerMeL employs first-order gradient-based FOMAML networks for disease identification in skin lesion datasets to ensure the approach remains lightweight. The method involves model pruning during the meta training phase to retain only common essential part of the model. The deleted connections are utilized during the meta testing phase to learn specific features of a rare disease in a limited dataset regime. DerMeL works best for rare disease prediction because pruning removes redundant or less crucial connections and parameters. This results in a concise and compact model. This size reduction can notably reduce memory demands, enhancing the model’s efficiency for deployment on devices or platforms with limited resources. Moreover, reconnecting these deleted connections allows quick learning of rare disease conditions in the meta test phase. The results show an increase in few-shot classification scores by approximately 2
Medication recommendation aims to predict effective drug combinations for patients with complex conditions. Critically-ill patients often experience polypharmacy, which increases the risk of adverse drug–drug interactions (DDIs). In this paper, we use a BERT-enhanced variation of Bayesian Personalized Ranking (BPR) algorithm for providing Toxic-free Medication Recommendations. Our model, denoted as ToxicFreeMed, integrates patients’ electronic discharge notes, drug descriptions, and the DDI knowledge graph to recommend both accurate and safe drug combinations. ToxicFreeMed leverages pretrained discharge notes embeddings-capturing patient’s disease, therapy, and medication context-and pretrained drug descriptions embeddings within a multitask learning framework that jointly optimizes ranking accuracy and a toxicity-weighted DDI loss. We have evaluated ToxicFreeMed on the MIMIC-III dataset augmented with DrugBank knowledge and show that it effectively balances accuracy and safety, outperforming other strong recommendation algorithms while reducing DDIs. Our results indicate that ToxicFreeMed is a promising approach that has the potential to assist clinicians in identifying therapeutically appropriate and pharmacologically safer treatment options.
The application of deep learning technology in medical text parsing has continuously progressed. Achieving high-quality extraction of complex entity relationships in medical texts has become an important research direction. Aiming at the problems of lengthy paragraphs and complex sentences in the field of psychological medicine, and the limited feature extraction ability of existing neural network relation extraction models, a psychological medicine entity relation extraction model MMF-RE, with multi-level and multi-unit gated convolution feature enhancement is proposed. The proposed MFE-BERT combines and outputs all the internal encoder layer features based on the pre-trained model to improve the semantic representation ability of the feature vector. At the same time, a multi-unit gated convolutional network is constructed, which can effectively extract multi-granularity local features to perceive long entities. Finally, the FNNAttention mechanism is applied to the model to strengthen the word-level relationship through the forward neural network function. The experimental results show that in the self-built psychological medicine dataset and the biomedical public dataset, the F1 value of the psychological medicine entity relation extraction of the MMF-RE composite neural network model reaches 88.49
Early detection of cognitive impairment remains a critical public health challenge. While biomarkers such as neuroimaging and cerebrospinal fluid analyses offer high sensitivity, their limited accessibility hampers widespread screening, especially in underserved settings. Speech-based markers have emerged as promising, noninvasive indicators of cognitive decline. To develop and validate SpeechDETECT, an end-to-end speech-processing pipeline that captures fine-grained acoustic and temporal markers of cognitive impairment and provides interpretable outputs suitable for large-scale screening. SpeechDETECT comprises six modules: (1) noise reduction / amplitude normalization; (2) an eight-domain voice-analysis framework (e.g., frequency parameters, speech fluency); (3) 50 ms segment-level feature extraction; (4) feature visualization; (5) dimensionality reduction / selection (Joint Mutual Information Maximization, LassoNet, PCA); and (6) classifier training with SHapley Additive exPlanations (SHAP). Performance was benchmarked against six acoustic toolkits (e.g., GeMAPS) on two English datasets: the DementiaBank Pitt corpus (train = 166, test = 71) with single cookie-theft picture description task and NIA PREPARE Phase 2 corpus (train = 1 064, test = 267) with multiple speech tasks. A Multi-Layer Perceptron trained on PCA-derived SpeechDETECT features achieved an F1-score = 0.81 https://github.com/SpeechCARE/SpeechDETECT-Toolkit for researchers and clinicians. A demo tutorial video showing pipeline usage is available at https://github.com/SpeechCARE/SpeechDETECT-Toolkit/blob/main/SpeechDETECT.mp4 .
The purpose of this work is to develop an Artificial Intelligence-based system capable of classifying individuals as either healthy or having a cardiac pathology (Myocardial Ischemia, Cardiomyopathy, Bundle Branch Block, Dysrhythmia, or Ventricular Hypertrophy) based on the processing of Electrocardiogram (ECG) signals. This system aims to serve as a non-invasive and efficient diagnostic tool. The ECG signals used in this study were obtained from a publicly available database. Signal processing is performed using Empirical Mode Decomposition (EMD). From the decomposition, statistical features are extracted from the modes obtained and structured into a feature vector that represents the signal. Classification is achieved using an optimized Neural Network (NN) model, which uses 2580 ECG signals to distinguish between healthy individuals and those with cardiac disease, considering cross-validation. After implementing the system, the mean accuracy achieved was 96.03
Polycystic Ovary Syndrome (PCOS) is a complex hormonal disorder that affects women of reproductive age globally. It is marked by hormonal imbalance, ovarian dysfunction, and metabolic problems, which may lead to serious health complications. Getting diagnosed early is crucial to detecting PCOS, with abdominal ultrasound being the most commonly used imaging modality. The conventional way of diagnosing PCOS relies on manual image assessment, which is time-consuming and can vary from person to person. To address existing limitations, this study proposes two automated PCOS classification frameworks based on deep feature fusion technique. Deep features were extracted from two complementary deep learning models, ResNet-18 and GoogLeNet, and fused to form a unified representation of ovarian characteristics. Then, the fused features were classified using two strategies: (i) an ensemble framework where AdaBoost was the meta-learner, and (ii) a lightweight custom classification head. The frameworks proposed in this study were evaluated against existing state-of-the-art approaches and demonstrated superior performance. The ensemble model achieved an accuracy of 97.36
Biomedical relation extraction (RE) aims to identify typed relations between biomedical entities from text. Most existing RE approaches rely on text classification frameworks, which struggle to fully exploit the generative capabilities of large language models (LLMs). Moreover, when LLMs are directly fine-tuned for biomedical RE, they often confuse semantically similar relation types, leading to the incorrect prediction of plausible but unsupported relations. To address these limitations, we propose BioRENLI, an LLM-based biomedical RE framework driven by preference-aligned natural language inference (NLI). BioRENLI reformulates RE as an NLI task by converting each candidate relation into a natural language hypothesis, prompting the model to predict either Entailment or Contradiction given the input context. To further mitigate label ambiguity, we augment supervised fine-tuning with Direct Preference Optimization (DPO). Using label-derived preference pairs, DPO encourages the model to prefer gold NLI decisions over non-gold alternatives, thereby reducing unsupported entailment predictions. Experimental results on the ChemProt and DDI datasets show that BioRENLI consistently outperforms supervised fine-tuned LLM variants in both full-supervision and low-resource settings, while remaining competitive with strong BERT-based baselines, highlighting the effectiveness of enhancing biomedical RE with preference-aligned NLI.
Alzheimer's disease (AD) is a progressive neurological disorder impacting a significant segment of the global population. Magnetic resonance imaging (MRI) is used to visualize brain structures and detect changes associated with AD. Early detection remains a major challenge, particularly for multiclass classification of disease severity. This paper presents an integrated hybrid framework for AD diagnosis combining Variational Mode Decomposition, fuzzy clustering, and multi-kernel learning. The proposed method comprises: (i) Variational Mode and Non-convex Optimized pre-processing for noise removal, (ii) Fuzzy Relevance Vector Machine-based segmentation to identify regions of interest, and (iii) Multi-Kernel SVM based clustering for classifying mild, moderate, and non-demented cases. The framework is validated on the Alzheimer's Disease Multiclass Images Dataset. Performance is evaluated using peak signal-to-noise ratio (PSNR), segmentation accuracy, training time, and precision. The results demonstrate that systematic integration of these established techniques achieves competitive performance, with average PSNR of 28.45 dB, segmentation accuracy of 91
The key to optimizing insulin administration and simplifying the management of Type 1 diabetes (T1D) lies in accurately predicting future blood glucose (BG) levels. Consistently predicting BG levels is a challenging goal due to interindividual biological variability, data quality issues, and the inherent variability of glucose metabolism. The study aims to predict BG levels across different time horizons by analyzing multimodal data from the BrisT1D and OhioT1DM datasets, comprising CGM measurements, insulin pump data, smartwatch activity data, and dietary carbohydrate data. The purpose of this research was to develop a robust time series model that could handle noise and heterogeneous medical data and that could contribute to clinical decision making for patients with T1D. A variety of time series transformer models were applied, and the best model was AutoBiGluNet, which is a hybrid deep learning model that uses Autoformer and BiLSTM networks to capture global patterns and temporal dependencies. Data were preprocessed by replacing missing values for time series features through linear interpolation and using zero imputation for other numeric values. AutoBiGluNet produced the best performance on BrisT1D, achieving an RMSE of 0.0674 ± 0.0006, MAE of 0.0411 ± 0.0004, and R2 of 0.9523 ± 0.0003 across five independent runs. On the external OhioT1DM dataset, the model also showed good generalizability, achieving RMSE of 0.88, MAE of 0.52, and R2 of 0.93 at the 30-minute prediction horizon. The model demonstrated strong predictive performance and favorable clinical error-grid results, suggesting potential for future decision-support applications. However, prospective clinical validation is required before considering integration into closed-loop insulin delivery systems.
Accurate reconstruction of the maxillomandibular relationship is essential for successful prosthetic rehabilitation in digital dentistry. Conventional registration techniques are severely constrained by the absence of direct overlap between the upper and lower dental arches and by their sensitivity to common clinical occlusal variations. In this paper, we introduce MAGIC-OR (Multi-dimensional Alignment with Geometric Intraoral Constraints for Occlusal Reconstruction), a framework that tightly couples multi-view intraoral scanning with geometric constraint optimization. The core of the method is a dual-buccal scanning strategy that exploits the buccal surfaces as a reliable registration bridge; an automatic selection mechanism identifies the optimal transformation among the bilateral scans, thereby coping effectively with occlusal asymmetry and unilateral data degradation. Building on this initial alignment, a geometry-constrained refinement engine with an enhanced safety metric (jointly penalizing penetration frequency, mean depth, and maximum depth) simultaneously evaluates biomechanical features-centroid distance, contact density, normal vector alignment, and gap distribution-to systematically suppress inter-arch penetration while promoting occlusal contact uniformity and stability. Comprehensive weight sensitivity analysis and ablation studies are conducted to quantify the contribution of each geometric constraint. Experiments on 155 clinical cases confirm that MAGIC-OR offers a notable advance in maxillomandibular registration, directly tackling the fundamental difficulty of limited inter-arch overlap through its geometric, dual-buccal optimization paradigm.
Functional magnetic resonance imaging (fMRI) provides a crucial window for understanding brain functional connectivity (FC) in psychiatric disorders, yet its complex spatiotemporal dynamics pose substantial challenges for modeling. Existing methods often rely on static FC, making it difficult to capture the dynamic plasticity of brain, while generally ignoring structural differences across functional networks or discarding informative weak connections due to excessive sparsification. Here, we propose SPSGL, a biologically inspired deep learning framework designed to construct novel brain connectivity patterns from fMRI signals. SPSGL transforms voxel-wise time series into frequency-domain, feature-driven functional brain graphs and employs a biologically inspired gated edge-update mechanism to capture dynamic changes in connectivity strength. On this basis, core functional networks and whole-brain patterns are mapped as structural priors to explicitly guide multi-head attention in forming complementary subspace foci that emphasize neurobiologically meaningful connections. Further combined with Orthonormal Clustering Readout (OCRead), our model achieves adaptive learning of multi-scale brain graph representations and functional parcellations. Across five psychiatry-related computational tasks, SPSGL demonstrates superior performance compared with existing approaches. Moreover, it identifies task-relevant functional connections and hub regions associated with aberrant coupling among the default mode, sensorimotor, and subcortical networks, highlighting potential neuroimaging biomarkers and uncovering shared brain network factors shared across diverse psychiatric conditions. Overall, SPSGL provides a unified, interpretable, and high-performing framework for fMRI-based brain connectivity analysis, advancing mechanistic understanding and potential clinical translation in mental health research. Our code is publicly available on https://github.com/zhaoqi106/SPSGL .
Major Depressive Disorder (MDD) represents a multifaceted and widespread mental health condition marked by substantial alterations in brain connectivity and neural dynamics, causing physiological stress. Accurate diagnosis via electroencephalography (EEG) is challenging due to the complexities in capturing both linear and nonlinear features of brain activity that vary over time and among individuals. Current EEG-based methods for MDD detection inadequately integrate spatial structure and temporal dependencies, often emphasising spatial proximity without functional relevance or using feature extraction techniques that neglect essential brain dynamics. To address these issues, we propose a novel advanced deep learning model: a 4D Graph Temporal Network (4D-GTNet) algorithm that works on a 4D graph brain network for MDD detection. The proposed methodology includes key innovations: A comprehensive 4D feature cube (A structured representation of EEG data combining spatial (channel), frequency (band), time (temporal slice), and statistical features. Each sample is modeled as a tensor or cube) from EEG data slices including linear and nonlinear features (statistical, temporal, and frequency) is extracted. An advanced nearest-neighbor-based method is employed to identify the functionally relevant spatial relationships among EEG channels. The integration of Graph Convolution and Gate recurrent Unit for capturing the spatiotemporal features is proposed with a channel-wise Max-Pooling module-Graph Temporal Max Pooling (GTMP) to retain critical channel-specific information by identifying the most informative channels. The experiment is conducted on data for 490 slices, 232 for MDD and 248 for Healthy Control (HC)) achieving 83.67% of accuracy. Experiments show that our proposed model outperforms the state-of-the-art models. Our method captures the complex dynamics of EEG signals effectively, enhances the understanding of brain connectivity patterns linked to MDD, and facilitates more accurate depression diagnoses.
Insomnia is a common sleep disorder, that causes difficulty in sleeping, staying asleep, or having non-restorative sleep. It often leads to daytime fatigue and impacts individuals’ well-being and daily functioning. Effective detection of insomnia is crucial for proper diagnosis and treatment planning. This paper proposes an accurate noninvasive detection of insomnia using photoplethysmography (PPG) signals. PPG signals offer a convenient and accessible method for continuous sleep monitoring without the need for specialized equipment. This study proposes PLETHSOMNet an automated insomnia detection using PPG and several deep learning (DL) models. To the best of our understanding, this is the first study to use PPG signals coupled with DL techniques for insomnia detection automatically. The proposed approach utilizes PPG signals extracted from the Cyclic Alternating Pattern (CAP) sleep database. Several deep-learning architectures were explored to classify individuals with insomnia and healthy sleepers. Model performance was evaluated based on different segment lengths of PPG signals, specifically 2 -second and 30 -second segments, to assess the system’s adaptability. For detecting insomnia automatically the proposed model has achieved the classification accuracy of 95 .89
For online medical consultation platforms, it is essential to recommend physicians with clinical expertise and interpersonal skills to improve patient satisfaction and clinical outcomes. Current physician recommendation is focused primarily on physicians' clinical qualifications and historical interactions, while largely overlooking the significant impact of interpersonal skills on effective patient interactions. To address this gap, we propose a novel framework called Self-supervised Fusion of Clinical Expertise and Interpersonal Skills for Enhanced Physician Recommendation. This framework jointly models physicians' clinical expertise and interpersonal skills through self-supervised learning. Specifically, we have designed a multi-attribute fusion paradigm that integrates three key physician attributes: (1) clinical expertise derived from physicians' profiles and historical dialogues, (2) interpersonal skills, including empathy, clarity, and responsiveness, measured through role-based emotional analysis of patient-physician interactions, and (3) patient evaluations that reflect their subjective experiences. Our self-supervised learning strategy aligns these diverse attributes into a unified representation space, automatically capturing the intrinsic correlations between physicians' expertise, interpersonal patterns, and patient feedback. Furthermore, the transformer-based encoder utilizes enhanced fusion representations and patient representations to assess the suitability of a physician to handle a specific patient consultation. Experiments on a real-world medical dialogue dataset demonstrate that SF-CEisEPR+FL outperforms the MUL-ATT baseline (a representative dialogue-driven method), achieving absolute gains of +6% in Precision@1, +5% in MAP, and +0.4% in ERR@5.
Postoperative pain management remains a major challenge due to substantial interindividual variability, and integrating genetic biomarkers with artificial intelligence (AI) offers a promising approach to precision analgesia. In this study, we investigated stress-responsive transcription factor ATF3 polymorphisms within a machine learning–ready framework to predict postoperative analgesic requirements. In a prospective cohort of 167 adults undergoing abdominal surgery, homozygous carriers of ATF3 SNPs rs3122721 and rs3125293 showed significantly higher opioid consumption independent of subjective pain scores, establishing a robust genotype–phenotype association suitable for algorithmic modeling. To enable translation, we developed a structured dataset architecture that supports real-time predictive analytics, allowing genetic profiles to serve as input features for deep learning models capable of forecasting high-risk patients and guiding personalized therapeutic strategies. These findings revealed persistent genotype-dependent opioid requirements over 72 h, providing a biological basis for clinical decision support systems that can dynamically adjust PCA protocols and dosing recommendations while incorporating safeguards to minimize opioid-related adverse effects. Collectively, this work identifies ATF3 genotyping as a promising biomarker for AI-driven precision analgesia, bridging genomic insight with actionable clinical strategies, and highlights practical considerations for perioperative integration along with the ethical implications of preoperative genetic testing.
Syncope is a transient loss of consciousness and increased fall risk due to cerebral hypoperfusion, often triggered by prolonged upright posture. Clinically, differential diagnosis largely relies on clinical manifestations and haemodynamic responses observed during the head-up tilt test (HUTT), whereas prediction is typically based on electrocardiographic information acquired during HUTT. However, early identification remains challenging because prodromal symptoms are often nonspecific and HUTT is time-consuming and may cause patient discomfort. A novel detection framework, AReS-Syncope, was designed to facilitate early syncope prediction by integrating metrics of autonomic imbalance and repolarization–energospectral alterations. Two feature sets were established: Autonomic Imbalance Features (AIF), capturing RR-interval prolongation and irregular variability preceding syncope and Repolarization–Energospectral Shift Features (RES), quantifying reduced repolarization reserve and a shift in spectral energy toward lower frequencies during postural stress and autonomic dysregulation. The dimensionally reduced AIF and RES sets were fed into a support vector machine classifier, and stability was evaluated across prediction windows of varying durations. The AReS-Syncope algorithm achieved early prediction of syncope using ECG signals alone, with a prediction horizon of 80 s before syncope. Using patient-level nested five-fold cross-validation, the model achieved an AUC of 91.42
Objective:This study endeavors to develop an intelligent diagnosis and treatment model for Traditional Chinese Medicine (TCM) syndrome differentiation and treatment, characterized by robust reasoning capabilities and exceptional reliability. This is achieved by harnessing clinical case data, employing knowledge distillation methodologies, and integrating Direct Policy Optimization (DPO) reinforcement learning techniques. Methods:GPT-4o was employed as the teacher model to perform knowledge distillation on TCM clinical case data, thereby generating a high-quality instruction dataset for syndrome differentiation and treatment. The distilled data was subsequently fine-tuned by Low-Rank Adaptation (LoRA) method grounded in the Qwen2.5-7B model to improve its abilities of diagnosis and treatment, together with competencies of personalized syndrome differentiation and treatment. Furthermore, an additional dataset of clinical cases was leveraged to emulate the diagnosis and treatment preferences of TCM experts, with the DPO reinforcement learning technique being applied for ongoing refinement and enhancement. Results:The recommended knowledge distillation approach maintained prescription recommendation performance comparable to direct training upon the reinforcement of the model's interpretability and generalization on external data. Aside from that, the DPO method ameliorated the stability of prediction results on the external dataset. Conclusion:The TCM Large Language Model constructed with knowledge distillation and reinforcement learning strategies effectively enhances diagnosis and treatment reasoning and personalized syndrome differentiation and treatment competencies. This approach provides new research directions and technical support for intelligent TCM clinical decision-making.
Under the “Internet Plus” model, this study aims to develop a proof‑of‑concept intelligent decision‑support system for elderly disease risk assessment, with a specific focus on Alzheimer’s Disease (AD). The primary objective is to evaluate whether blood‑based transcriptomic biomarkers can be translated into an interpretable and scalable risk stratification framework suitable for home‑care and community health settings. Whole‑blood transcriptomic and associated clinical data were obtained from a publicly available microarray dataset comprising 329 individuals (144 AD, 104 MCI, and 81 cognitively normal controls). Transcriptomic features were used exclusively for model training, while clinical variables—including cognitive scores, frailty index, depression scores, and medication adherence—were used post hoc for risk annotation and alert simulation. Rigorous preprocessing and quality control were applied, followed by differential gene expression analysis using the limma framework and biologically informed biomarker selection. A Random Forest classifier trained on selected transcriptomic biomarkers achieved an accuracy of 91.2
Cardiovascular diseases (CVD) are among the leading causes of mortality worldwide due to genetic predisposition and lifestyle factors. Proper diagnosis of cardiovascular diseases is crucial to provide early-stage treatments. Conventional diagnostic methods such as stress tests, electrocardiograms, and echocardiography detect valuable insights into rhythm abnormalities, structural anomalies, or other cardiovascular conditions. However, their reliability heavily depends on human expertise, and they may not always detect early-stage signs of disease. In recent years, Machine Learning (ML) models have emerged as alternative diagnosis tools, capable of identifying CVD with higher accuracy. ML enables automated and precise detection based on data relationships, capturing hidden, complex patterns that are not apparent through traditional diagnostics. Most ML approaches employ supervised learning, which requires labeled data that are not always available in medical records. Under such circumstances, unsupervised learning has been explored as a suitable alternative. In this paper, a hybrid unsupervised approach combines the neural network structure of Self-Organizing Maps (SOM) with the dimensionality reduction technique of Principal Component Analysis (PCA) for unsupervised analysis for clustering CVD across different severity levels. Considering a data compression mechanism, the synergy among these methods leverages the ability to map unsupervised complex, high-dimensional data into lower-dimensional space. The proposed approach significantly improves the detection of hidden structures within large, high-dimensional medical cardiovascular datasets, providing insights into cardiovascular risk factors and improving the overall diagnostic process. Experimental evaluation on the UCI Cleveland Heart Disease dataset shows that the proposed PCA-SOM model achieves a Silhouette score of 0.94 (train) and 0.79 (test), and a Davies-Bouldin index of 0.08 (train) and 0.16 (test), outperforming baseline clustering methods such as K-means, hierarchical clustering, Gaussian Mixture and Spectral clustering highlighting its potential for supporting CVD detection.
