Background & Aims Monitoring and managing complications after liver transplantation (LT) are crucial for ensuring graft and patient survival. This study aimed to investigate the association between liver stiffness measurement (LSM) and spleen stiffness measurement (SSM) by sound touch elastography (STE) with post-LT complications, and to develop a predictive post-LT complications (PLTC)-nomogram. Methods We conducted a retrospective study of patients who received LT between January 2019 and March 2024. After collecting clinical parameters and STE measurements, we constructed a prediction model using univariate and multivariate logistic regression, visualized as a nomogram. Its performance was evaluated with the area under the receiver operating characteristic curve (AUC), precision-recall (PR) curve, calibration curve, and decision curve analysis (DCA). The nomogram’s performance was also compared with LSM, SSM, aspartate aminotransferase-to-platelet ratio index (APRI), and fibrosis-4 index (FIB-4). Results A total of 113 recipients were included in the study. Post-LT complications occurred in 41 (36.3%) recipients, including rejection, vascular, biliary, renal, and malignant complications. Multivariate logistic regression analysis identified five factors independently associated with post-LT complications: LSM (odds ratio [OR], 2.64; 95% confidence interval [CI], 1.60–4.35), alkaline phosphatase (OR, 1.02; 95% CI, 1.01–1.04), total bilirubin (OR, 1.08; 95% CI, 1.01–1.15), creatinine (OR, 1.04; 95% CI, 1.02–1.07), and white blood cell count (OR, 0.42; 95% CI, 0.25–0.72). These five factors were used to develop the PLTC-nomogram. The nomogram demonstrated excellent performance with an AUC of 0.968 (95% CI, 0.940–0.996), outperforming LSM (AUC = 0.846), SSM (AUC = 0.676), APRI (AUC = 0.758) and FIB-4 (AUC = 0.800). Area under the PR curve (0.951), calibration curve, and DCA further confirmed that the PLTC-nomogram provided robust diagnostic performance. The PLTC nomogram is available via an online platform (https://AYGY-PLTC.shinyapps.io/dynnomapp/). Conclusions The PLTC-nomogram incorporating clinical parameters and LSM by STE offers a reliable and noninvasive method for predicting post-LT complications.

Evaluating a classifier's performance is critical for its successful application. This paper explores various metrics used for binary classification tasks, highlighting their strengths and limitations. Simple threshold metrics, such as Accuracy and Sensitivity, are efficient for binary data and a single cutoff point. However, their reliance on a single threshold and sensitivity to imbalanced data can be drawbacks. For more robust evaluation, ranking metrics such as Receiver Operating Characteristic (ROC) and Precision-Recall (PR) curves provide a threshold-agnostic approach, enabling comparison across different cutoff points. Additionally, probabilistic metrics like Brier Score and Log Loss assess the model's ability to predict class probabilities. The choice of metric depends on the specific classification problem and the characteristics of the data. When dealing with imbalanced data or complex decision-making processes, using multiple metrics is recommended to gain a comprehensive understanding of the model's performance. This paper emphasises the importance of understanding metric limitations and of selecting appropriate metrics for a specific classification task. By doing so, researchers and practitioners can ensure a more accurate and informative evaluation of their models, ultimately leading to the development of reliable tools for various applications.

Precise lung region detection in chest radiographs is an essential preprocessing step for computer-aided diagnostics. This study presents a YOLO v4–based framework to automatically localize lung regions in posteroanterior (PA) chest X-rays. A subset of the CheXpert dataset, containing 456 manually annotated PA radiographs, was used. Anchor boxes were estimated via an Intersection-over-Union (IoU)–based clustering method, improving scale invariance and shape alignment over Euclidean metrics. Empirical evaluation showed that six anchor boxes achieved the best balance between mean IoU (0.883) and computational efficiency. The trained model was tested on 144 images, yielding Average Precision (AP) of 0.9043 for the lung_region class, which represents only the anatomical lung area and not any specific pathology. The precision–recall curve indicated high precision across most recall values, and the confusion matrix showed 124 true positives, 13 false positives, and 7 false negatives. These results demonstrate that YOLO v4 with optimized anchor box estimation enables accurate, efficient lung region localization, supporting automated radiology workflows.

Diabetes falls within the category of chronic diseases, and the issue of its prevention and control has always been a health concern for all mankind. This study constructs a prediction model for diabetes based on four machine learning methods: Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). In the distribution of important features in LR, the top five important features are selected. On this basis, a model is constructed through other algorithms, aiming to reduce the interference of other irrelevant features, and then compare the performance of models constructed by different machine learning algorithms. The dataset of this study contains 768 samples, covering 8 characteristics including metrics like the number of pregnancies and plasma glucose concentration. Finally, the accuracy rates of the models constructed by the four algorithms of LR, DT, RF, and XGBoost were 75.97%, 78.57%, 74.03% and 74.68% respectively. Combining the area under the receiver operating characteristic curve (AUC), as well as the precision rate, recall rate, and F1 score of the positive category, it is ultimately concluded that the predictive effect of DT is the best. However, to better integrate diabetes prediction models with practical applications, more data and resources are still needed to support them.

Bladder cancer represents a significant health problem worldwide, with it being a major cause of death and characterized by frequent recurrences. Effective treatment hinges on early and accurate diagnosis; however, traditional methods are invasive, time-consuming, and subjective. In this research, we propose a transparent deep learning model based on the YOLOv11 structure to not only enhance lesion detection but also give the visual support of the model’s predictions. Five versions of YOLOv11—nano, small, medium, large, and extra large—were trained and tested by us on a comprehensive dataset of hematoxylin and eosin-stained histopathology slides with the inflammation, urothelial cell carcinoma (UCC), and invalid tissue categories. The YOLOv11-large variant turned out to be the best-performing model at the forefront of technology, with an accuracy of 97.09%, precision and recall of 95.47% each, and balanced accuracy of 96.60%. Besides the precision–recall curves (AUPRC: inflammation = 0.935, invalid = 0.852, UCC = 0.958), ROC-AUC curves (overall AUC = 0.972) and risk–coverage analysis (AUC = 0.994) were also used for detailed assessment of the model to confirm its steadiness and trustworthiness. The confusion matrix displayed the highest true positive rates in all classes and a few misclassifications, which mainly happened between inflammation and invalid samples, indicating a possible morphological overlap. Moreover, as supported by a low Expected Calibration Error (ECE), the model was in great calibration. YOLOv11 reaches higher performance while still being computationally efficient by incorporating advanced architectural features like the C3k2 block and C2PSA spatial attention module. This is a step towards the realization of the AI-assisted bladder cancer diagnostic system that is not only reliable and transparent but also scalable, presented in this work.

BackgroundEmergency Department (ED) revisits represent a critical issue in emergency medicine. Identifying high-risk revisit cases (revisits with intensive care unit admissions, cardiac arrest, or requiring emergency surgery) is particularly important. While prior studies have explored machine learning models for ED revisit prediction, few deep learning approaches exist, and dynamic features remain underutilized.MethodsWe used data from National Taiwan University Hospital (NTUH), incorporating both static (e.g., age, sex, triage) and dynamic (vital signs) features. A preprocessing strategy was developed to handle temporal irregularities. We proposed a hybrid deep learning model combining Temporal Convolutional Network (TCN) and feature tokenizer (FT)-Transformer to integrate static and short-term dynamic information.ResultsWe evaluated our model on NTUH 2016–2019 data, achieving the area under the receiver operating characteristic curve (AUROC) of 0.8453 and the area under precision recall curve (AUPRC) of 0.0935 for high-risk revisits (base rate = 0.01), and AUROC of 0.7250 and AUPRC of 0.2005 for general revisits (base rate = 0.042). The model maintained robust performance when validated on 2020–2022 data. Compared to the static-only logistic regression baseline, our model improved AUPRC from 0.0288 to 0.0935 and precision from 0.0281 to 0.0428.ConclusionOur model significantly outperformed the static-only baseline. It demonstrates the effectiveness of multimodal clinical data fusion in improving ED revisit prediction and supporting clinical decision-making.

This study validates an automated seizure detection algorithm for multi-channel neonatal EEG, adapting a previously published method to a dataset with fewer electrodes. The Python-based implementation, SDApy, was applied to EEG recordings from 23 neonates to classify seizure and non-seizure epochs using a support vector machine trained on an independent dataset. The algorithm employs time- and frequency-domain features and maintains high generalization across different recording setups, achieving robust performance despite using only nine electrodes instead of nineteen. Evaluation metrics, including F1 scores and precision—recall curves, confirmed strong agreement between algorithm predictions and expert annotations for most patients. SDApy’s open-source implementation enhances accessibility compared with earlier MatLab versions, offering a transparent and cost-effective approach to clinical EEG analysis. The pipeline can operate with labels from a single expert, supports data pre-labeling for deep learning, and integrates well into neonatal intensive care unit monitoring workflows. Overall, SDApy demonstrates reliable adaptation to reduced-channel EEG and shows potential for real-time seizure detection, personalized threshold optimization, and integration into multimodal neurophysiological monitoring systems.

Traffic safety agencies seek predictive models that identify high-risk crashes before traffic accidents occur, but realworld crash data are highly imbalanced, rendering the overall accuracy of machine learning models a misleading metric for predictive performance. This study investigates a recall-first approach to crash-risk classification in Virginia, arguing that missing a dangerous event carries a much higher cost than issuing an extra alert based on the data provided by Virginia Department of Transportation (VDOT). The modeling framed high-risk identification as a binary classification task and used standard tools, Logistic Regression and Random Forest, augmented with imbalance-aware training. Evaluation emphasized recall and precision, utilizing precision–recall curves and confusion matrices. Across experiments, recall-oriented training and thresholding consistently improved detection of high-risk cases relative to accuracy-optimized baselines, with expected trade-offs resulting in lower precision and overall accuracy. This paper further translates operational priorities into simple threshold policies, showing how agencies can tune models based on recall to align with resource constraints while minimizing costly misses. In conclusion, for safety-critical applications, recall should be the primary metric of success; furthermore, straightforward imbalance treatments—without complex model additions—can realign model performance with public-safety goals.

Background and Objectives: Type 2 diabetes mellitus (T2DM) affects over 537 million adults worldwide and disproportionately burdens low- and middle-income countries, where diagnostic resources are limited. Predictive models trained in one population often fail to generalize across regions due to shifts in feature distributions and measurement practices, hindering scalable screening efforts. Materials and Methods: We evaluated a few-shot domain adaptation framework using a simple multilayer perceptron with four shared clinical features (age, body mass index, mean arterial pressure, and plasma glucose) across three adult cohorts: Bangladesh (n = 5288), Iraq (n = 662), and the Pima Indian dataset (n = 768). For each of the six source-target pairs, we pre-trained on the source cohort and then fine-tuned on 1, 5, 10, and 20% of the labeled target examples, reserving the remaining for testing; a final 20% few-shot version was compared with threshold tuning. Discrimination and calibration performance metrics were used before and after adaptation. SHAP explainability analyses quantified shifts in feature importance and decision thresholds. Results: Several source → target transfers produced zero true positives under the strict source-only baseline at a fixed 0.5 decision threshold (e.g., Bangladesh → Pima F1 = 0.00, 0/268 diabetics detected). Few-shot fine-tuning restored non-zero recall in all such cases, with F1 improvements up to +0.63 and precision–recall gains in every zero-baseline transfer. In directions with moderate baseline performance (e.g., Bangladesh → Iraq, Iraq → Pima, Pima → Iraq), 20% few-shot adaptation with threshold tuning improved AUROC by +0.01 to +0.14 and accuracy by +4 to +17 percentage points while reducing Brier scores by up to 0.14 and ECE by approximately 30–80% (suggesting improved calibration). All but one transfer (Iraq → Bangladesh) demonstrated statistically significant improvement by McNemar’s test (p < 0.001). SHAP analyses revealed population-specific threshold shifts: glucose inflection points ranged from ~120 mg/dL in Pima to ~150 mg/dL in Iraq, and the importance of BMI rose in Pima-targeted adaptations. Conclusions: Leveraging as few as 5–20% of local labels, few-shot domain adaptation enhances cross-population T2DM risk prediction using only routinely available features. This scalable, interpretable approach can democratize preventive screening in diverse, resource-constrained settings.

Quantitative vascular analysis involves the measurements of arterial tortuosity and branch angle in a region of interest in cerebral arteries to assess vascular risks associated with cerebral aneurysm. The measurements themselves are not a simple process since they are made on the three-dimensional (3D) structures of the arteries. The aim of this study was to develop a deep convolutional neural network (CNN) model to predict a probability score of aneurysm without direct measurements of the artery’s geometry. A total of 204 subjects’ image data were considered. In all, 585 gray-scale three-dimensional (3D) patches with the bifurcations near the center of the patches were extracted and labeled as either an aneurysm or a non-aneurysm class. Three-dimensional CNN architectures were developed and validated for the binary classification of the 3D patches. Accuracy, precision, recall, F1-score, receiver operating characteristics area under the curve (ROC-AUC), and precision recall AUC (PR-AUC) were calculated for test data. Deep learning predictions were compared with vessel geometry measurements. Deep learning probability scores were dichotomized into high-score and low-score groups. For both groups, bifurcation angles and sum-of-angles-metric (SOAM) were calculated and compared. ResNetV2_18 with translation as data augmentation achieved the highest mean ROC-AUC (0.735) and PR-AUC (0.472). The independent t-test indicated that for the bifurcation angle sum feature there was a statistically significant difference (t = −2.280, p-value < 0.05) between the low-score and the high-score groups. In conclusion, we have demonstrated a deep learning-based approach to the prediction of aneurysmal risks in the bifurcation regions of interest. Deep learning predictions were associated with vessel geometry measurements. This suggests that deep learning on 3D patches centered around the bifurcations has the potential to screen bifurcations with a high aneurysm risk.

Sexual minority adults are at elevated risk for eating disorders (EDs), yet existing screening tools have rarely been validated in this population. Most ED screening instruments have been validated in predominately cisgender, heterosexual female samples limiting their generalizability to populations with different symptom patterns. Validation studies in cisgender sexual minority (SM) adults are critical to improving detection and addressing disparities in ED identification. The present study evaluated the psychometric performance of the Brief Assessment of Stress and Eating (BASE), a validated 10-item screening tool that assesses DSM-5-aligned eating disorder symptoms and subclinical dysregulated eating behaviors, in a national sample of cisgender gay men and lesbian women. Participants were 1,498 cisgender SM adults (61.7% gay men, 38.3% lesbian women) recruited from The PRIDE Study, a U.S.-based longitudinal cohort of sexual and gender minority adults. Respondents completed the BASE, SCOFF questionnaire, and the Eating Disorder Diagnostic Scale-5 (EDDS-5) which we used to derive probable DSM-5 eating disorder (probable ED) status. Receiver operating characteristic (ROC) and precision-recall (PR) curve analyses were conducted to evaluate classification accuracy and identify optimal thresholds. Both the BASE and SCOFF performed significantly above chance in detecting EDDS-5-derived probable EDs. Among gay men, the BASE (AUC: ROC = 0.785, PRC = 0.702) outperformed the SCOFF (ROC = 0.744, PRC = 0.630). In lesbian women, the two screeners performed similarly (BASE AUC = 0.807; SCOFF AUC = 0.806). Optimal BASE thresholds varied by group with higher sensitivity at lower cutoffs (e.g., ≥ 7). The BASE provides a reliable, efficient alternative to traditional instruments for screening eating disorders in sexual minority adults, with good performance for identifying EDDS-5-derived probable EDs. Findings support the BASE as a reliable and valid screening tool for use with cisgender SM adults in community, healthcare, and research contexts.

Integrating WGCNA and machine learning to distinguish active pulmonary tuberculosis from latent tuberculosis infection based on neutrophil extracellular trap-related genes.

Substance use among youth is a significant public health issue, particularly in low resource settings in Sub-Saharan Africa (SSA), where it contributes to HIV transmission and poor engagement in HIV care. This study employs machine learning (ML) techniques to develop models for predicting problematic substance use (PSU) among youth living with HIV (YLHIV) in Uganda, aiming to identify important multilevel risk factors and compare predictive performance of ML algorithms. Utilizing a cross-sectional dataset of 200 YLHIV aged 18-24 in Uganda, we trained and evaluated six predictive models, through 10-fold cross validation. Model performance was assessed using area under receiver operating characteristic curve (AUROC), and precision recall curve (AUPRC). Subsequent feature importance analysis revealed key predictors of PSU. The random forest model achieved the best discriminative performance with an AUROC of 0.78 (0.01) and AUPRC of 0.75 (0.02). Key predictors of PSU spanned individual, interpersonal, and community dimensions including depression, sexual risk-taking behaviors, monthly income, adverse childhood experiences, family involvement in selling alcohol, friends enabling access to alcohol, exposure to community educational campaigns against alcohol, household size, and knowledge of alcohol effects on HIV treatment. Our findings highlight ML's potential in predicting PSU among YLHIV and provide insights to guide targeted interventions and support policy formulations mitigating PSU effects on HIV management.

Software requirement specifications (SRS) often contain repetitive, ambiguous, or inconsistent requirements, which drives up costs and delays project timelines. Because they mainly rely on syntactic similarity, traditional redundancy detection methods like TF-IDF and Word2Vec have trouble detecting semantic overlaps. This study proposes a semantic pruning framework that uses advanced NLP techniques, with a focus on transformer-based models like BERT, to find and eliminate superfluous requirements from SRS documents. Precision, recall, F1-score, and runtime were used as evaluation criteria to compare several methods, including CountVectorizer, TF-IDF, Word2Vec, and BERT. The findings show that deep learning models outperform conventional methods, which yield high precision but poor recall. Despite having a longer runtime, BERT outperformed Word2Vec with F1 = 0.87 and recall = 0.77. The outcomes show the effectiveness of transformer-based embeddings for re-dundancy detection and provide a scalable approach to improve SRS quality while reducing the amount of manual review effort

Accurate identification of advanced tumor stages is essential for timely clinical decision-making and personalized treatment planning. This study proposes an explainable ensemble learning framework for predicting advanced tumor stage using a dataset containing 10,000 samples with 18 clinical and radiological features. Four machine learning models, namely Logistic Regression, Naïve Bayes, AdaBoost, and LightGBM, were evaluated using stratified train–test splits along with standard performance metrics. LightGBM achieved the highest performance, with an accuracy of 86.05% and an F1-score of 76.61%, outperforming linear and probabilistic classifiers. ROC–AUC and precision–recall analyses further confirmed the superior discriminative ability of ensemble methods. SHAP explainability techniques highlighted mitotic count, Ki-67 index, enhancement, and necrosis as the most influential predictors of advanced stage. The proposed framework demonstrates strong predictive capability and provides clinically interpretable insights, underscoring its potential as a decision-support tool in oncological diagnostics. Future work will involve external validation and integration of additional multimodal data to enhance generalizability.

Abstract Identifying the frequencies between drugs and side effects is crucial in drug development. As clinically validating potential side effects is financially expensive and time consuming, computational methods offer an appealing alternative for predicting candidate side effects. Since this prediction problem is inherently complex, many methods treated the problem as two tasks: association identification and frequency estimation. These tasks differ greatly in their objectives and learning dynamics, as association demands discrete classification, while frequency requires continuous regression. The previous multi-task based methods employ symmetric architectures that fail to address this divergence. To tackle this issue, we propose a novel method that employs a novel dual-task approach with asymmetric model architecture with a dedicated sub-network for each task. Our empirical study demonstrates that the proposed method achieves the best performance compared with the state-of-the-art approaches achieving 15.6% and 2.8% improvement in area under the precision–recall curve for classification, and achieving 2.2% and 31.8% reduction in mean absolute error for regression compared to the second-best method under warm-start and cold-start settings, respectively. Further experiments suggest there is a performance gain in using an asymmetric architecture than symmetric architectures.

Alzheimer’s Disease (AD), the most common neurodegenerative disease, is underdiagnosed and more prominent in underrepresented groups. We performed semi-supervised positive unlabeled learning (SSPUL) coupled with racial bias mitigation for equitable prediction of undiagnosed AD from diverse populations at UCLA Health using electronic health records. SSPUL achieved superior sensitivity (0.77–0.81) and area under the precision recall curve (AUCPR) (0.81–0.87) across non-Hispanic white, non-Hispanic African American, Hispanic Latino, and East Asian groups compared to supervised baseline models (sensitivity: 0.39–0.53; AUCPR: 0.3–0.7). SSPUL also exhibited superior fairness as evidenced by the lowest cumulative parity loss. We identified top shared and distinct features among labeled and unlabeled AD patients, including those that are neurological (e.g., memory loss) and non-neurological (e.g., decubitus ulcer). We validated our results using polygenic risk scores, which were higher in labeled and predicted positives than in predicted negatives among non-Hispanic white, Hispanic Latino, and East Asian groups (p < 0.001).

High mortality rates in maintenance hemodialysis (MHD) patients necessitate precise predictive tools. Existing models lack accuracy and ease of clinical access. This study focuses on constructing a precise and user-friendly machine learning-based mortality risk predictive model for MHD patients. A total of 601 MHD patients from Shantou Central Hospital were enrolled in this study. Clinical and laboratory data were meticulously gathered and assessed. Patients were divided randomly into Training (70%) and Test cohort (30%). Six types of machine learning algorithms based predictive models were constructed for prognostic prediction. The predictive accuracy of the model was evaluated by calculating the area under the receiver operating characteristic curve (AUROC) and the area under the precision recall curve (AUPRC). Additionally, an online predictive model application was developed for practical clinical application. The Training and Test cohort exhibited comparable demographic and clinical traits. Age, BMI, HGB, CH, AST, and serum albumin levels emerged as significant independent predictors of prognosis. The Extreme Gradient Boosting (XGBoost) based model predictive performance measures included with AUROC 0.831 and AUPRC 0.310 in the Test cohort. The XGBoost based model was selected as the definitive predictive tool and was made accessible via a web application. We successfully developed a machine learning-driven predictive model to predict the risk factors of MHD patients, which was then integrated into a user-friendly web application. This predictive tool could help to identify the high-risk factors of MHD patients in clinical practices.

Early detection of breast cancer by mammography scans is crucial for improving treatment outcomes. However, low image resolutions, size, and location of lesions in dense breast tissue prove to be challenges in mammography, underscoring the importance of accurate and efficient computer-aided diagnostic systems. This article introduces a novel classification framework that utilizes histogram of oriented gradients (HOG) as a feature extractor and principal component analysis (PCA) for dimensionality reduction. Classification is implemented using the persistent homology classification algorithm (PHCA), which leverages persistent homology (PH) to capture topological properties of mammography images. The framework was evaluated on 7,632 images from the INbreast dataset with an extensive use of grid-search cross-validation to optimize preprocessing parameters. Two optimal combinations of HOG parameters and scaler were identified, with the best configuration (16 × 16 pixels per cell, 3 × 3 cells per block, and Minmax scaler) achieving strong performance. Validating on the test set, PHCA achieved an overall accuracy, precision, recall, F1-score, and specificity of 97.31%, 96.86%, 97.09%, 96.97%, and 96.86%, respectively. Clinically, the high precision (98.23%) and high recall (97.75%) for malignant cases highlight PHCA’s sensitivity in identifying malignancies, ensuring that very few malignant cases go undetected with highly trustworthy predictions. These results are shown to be competitive with existing state-of-the-art models, even exceeding in some cases, while requiring lower computational cost than deep learning-based approaches. Although the proposed method trails advanced deep models by 3–4% in some metrics, it offers a computationally efficient alternative and a potential for deployment in large-scale screening systems, demonstrating the promise of topological data analysis for breast cancer classification.

Policymakers in small open economies need reliable signals of incipient private consumption downturns, yet traditional indicators are revised, noisy, and often arrive too late. This study develops a Romanian-specific early warning system that combines a time-varying parameter VAR with stochastic volatility and exogenous drivers (TVP-SV-VARX) with modern machine learning classifiers. The structural layer extracts regime-dependent anomalies in the macro-financial transmission to household demand, while the learning layer transforms these anomalies into calibrated probabilities of short-term consumption declines. A strictly time-based evaluation design with rolling blocks, purge and embargo periods, and rare-event metrics (precision–recall area under the curve, PR-AUC, and Brier score) underpins the assessment. The best-performing specification, a TVP-filtered random forest, attains a PR-AUC of 0.87, a ROC-AUC of 0.89, a median warning lead of one quarter, and no false positives at the chosen operating point. A sparse logistic calibration model improves probability reliability and supports transparent communication of risk bands. The time-varying anomaly layer is critical: ablation experiments that remove it lead to marked losses in discrimination and recall. For implementation, the paper proposes a three-tier WATCH–AMBER–RED scheme with conservative multi-signal confirmation and coverage gates, designed to balance lead time against the political cost of false alarms. The framework is explicitly predictive rather than causal and is tailored to data-poor environments, offering a practical blueprint for demand-side macroeconomic early warning in Romania and, by extension, other small open economies.