
Background: Colorectal cancer remains a major cause of cancer-related mortality worldwide, with poor outcomes in patients with advanced-stage disease. The c-Jun N-terminal kinase (JNK) pathway and hypoxia are important regulators of tumor progression, but their combined prognostic value in colorectal cancer has not been systematically evaluated using machine learning. Methods: Transcriptomic and clinical data from 346 patients with colorectal cancer were divided into training (n = 242) and testing (n = 104) cohorts. Pathway-specific model selection used nested cross-validation with five outer and five inner folds. Within each inner-training fold, minimum redundancy maximum relevance (mRMR) feature selection was repeated, and gene counts of 1–15 were evaluated across nine classifiers, retaining age and sex. An exploratory integrated analysis evaluated combinations selected from the resulting pathway-specific gene pool. An age-and-sex-only model served as a demographic baseline. Results: Results: Training-based selection identified random forest with eight JNK-related genes and histogram-based gradient boosting with 12 hypoxia-related genes. Their testing ROC-AUCs were 0.658 (95% CI: 0.538–0.774) and 0.600 (95% CI: 0.492–0.717), respectively; pooled outer out-of-fold ROC-AUCs were 0.549 and 0.536. The exploratory integrated analysis selected a common 12-gene set. Following parameter exploration and testing-based comparison, K-nearest neighbors achieved a testing ROC-AUC of 0.690 (95% CI: 0.584–0.786), compared with 0.483 (95% CI: 0.368–0.602) for the age-and-sex baseline. Conclusions: Machine learning analysis of JNK- and hypoxia-associated gene signatures suggests potential for predicting 2-year survival in colorectal cancer. The exploratory integrated findings warrant independent external validation, assessment of calibration, and comparison with clinicopathological models before clinical application.
Background: Biomedical Knowledge Graphs (KGs) such as PrimeKG and Hetionet offer vast structured repositories for clinical discovery; however, conventional Graph Neural Networks (GNNs) suffer from black-box opacity, while standard Retrieval-Augmented Generation (RAG) paradigms face unsubstantiated associations and mechanistic inconsistencies. This study aimed to develop BioAgent-RAG, a multi-agent orchestrated retrieval-augmented generation framework tailored for complex biomedical reasoning and transparent graph traversal to resolve this critical trade-off. Methods: The framework integrates autonomous sub-agents for query decomposition, path-based KG grounding, and critic-guided cross-verification to systematically enhance biological plausibility and minimize ungrounded inferences. It was evaluated through rigorous empirical comparisons against competitive baselines (RD2-KG, KGiA, RAG-LLMs, and TxGNN) on the PrimeKG and Hetionet benchmarks, measuring ROC-AUC, PR-AUC, MRR, Recall@10, Recall@50, Biomedical Factuality Score (BFS), and Path Reliability Score (PRS). Results: BioAgent-RAG substantially outperforms existing baselines, achieving on PrimeKG an ROC-AUC of 0.952, PR-AUC of 0.814, MRR of 0.637, and superior Recall@10 and Recall@50 of 0.715 and 0.868, respectively (p < 0.01). Moreover, it yields a Biomedical Factuality Score (BFS) of 0.887 on PrimeKG (0.894 on Hetionet) and a Path Reliability Score (PRS) of 0.841 (0.862 on Hetionet), demonstrating strong preliminary correlation (0.912) with domain-expert evaluations. Conclusions: These findings demonstrate that BioAgent-RAG serves as an actionable, clinically explainable computational screening platform capable of de-risking wet-lab pipelines by prioritizing high-confidence repurposing candidates and delivering verifiable multi-hop mechanistic hypotheses directly interpretable by pharmacologists.
Background/Objectives: Machine learning-based survival analysis methods often outperform classical approaches but entail greater methodological complexity and computational cost. This study proposes a unified three-component framework for selecting optimal survival modeling strategies by jointly considering data characteristics and research objectives. Methods: The framework comprises variable selection, classical and tree-based machine learning survival models, and performance evaluation using both time-aggregated and time-resolved metrics. It is validated on two distinct datasets of heart failure and leprosy, representing homogeneous and heterogeneous survival structures. Results: It is found that the optimal machine learning approach and variable selection are data dependent. For the heart failure dataset, a bagging survival tree achieves optimal performance, using only six predictors. For the leprosy dataset, the full predictor model yields the best performance, while the computationally efficient random survival forest performs comparably to the more complex bagging approach. Beyond model selection, the framework explicitly incorporates secondary objectives, including parsimony and computational efficiency, and supports the derivation of interpretable decision pathways for risk stratification. Conclusions: Heart failure patients with an ejection fraction below 20% should be considered as the highest risk group, whereas leprosy patients with symptom duration under one month can be regarded as the lowest risk category. In short-term prediction settings (less than 50 days/one year for heart failure/leprosy), the classical Cox proportional hazards model performs comparably to machine learning methods and may be chosen as the preferred method for simplicity. Overall, the proposed multi-component framework enables context-aware model selection by balancing predictive performance, model complexity, and temporal considerations.
Background: Automated QRS notch detection is challenging because QRS amplitude, morphology, and notch prevalence vary substantially across the 12 ECG leads. This study aimed to develop a lead-specific deep learning framework that explicitly accounts for these lead-dependent differences. Methods: The framework combined lead-targeted modeling with anatomically related context leads, lead-specific parameter optimization, and lead-matched architectures. Eleven leads were modeled using the target lead and related context leads, whereas aVR was modeled independently because of its distinct reciprocal orientation. The framework was evaluated using 21,799 PTB-XL recordings with PTB-XL+ notch annotations and patient-wise five-fold cross-validation. Results: The proposed framework achieved a mean lead-level F1-score of 84.20%, improving performance by approximately 13.29 percentage points over the best pooled baseline across the 12 leads. Ablation analyses showed improvements from anatomical context input (+4.68 percentage points) and lead-specific parameter optimization (+4.77 percentage points), while the lead-matched architectures also consistently outperformed the baseline architectures. Conclusions: These findings indicate that QRS notch detection benefits from explicitly modeling lead-level anatomical and statistical differences rather than applying a uniform model across all ECG leads.
Background: Short-term blood glucose prediction usually relies on continuous glucose monitoring (CGM) data sampled every 5 min, which are often unavailable in outpatient care. This study developed a computational workflow that transforms event-driven records into physiologically informed features for 30 min glucose prediction in insulin-dependent diabetes mellitus (IDDM). Methods: Records of 300 virtual patients (30 days, 30 min resolution), simulated from real outpatient glucose measurements and including glucose readings, insulin doses, and carbohydrate intake, were converted into a seven-dimensional representation. Pharmacokinetic decay functions encoded active rapid- and long-acting insulin and carbohydrate absorption, plus the glucose rate of change and a circular time-of-day encoding. The representation was evaluated with artificial neural network (ANN), support vector machine (SVM), random forest (RF), and gradient boosting (XGBoost) regression on 345,360 training and 86,340 test samples. Results: All algorithms reached a root mean square error (RMSE) of 12.04–12.28 mg/dL and R2 of 0.9845–0.9851, and Clarke Error Grid Analysis placed 93.7–93.9% of predictions in Zone A. Conclusions: The convergence of the four algorithms suggests that predictive performance depends mainly on the feature representation. The workflow offers a reproducible and interpretable basis for glucose prediction without high-frequency CGM data, pending validation on real patient records.