IEEE Access

I.AO_SCPLOWBSTRACTC_SCPLOWCoronary Artery Disease (CAD) is a leading cause of cardiovascular-related mortality and affects 20.5 million people in the United States and approximately 315 million people worldwide in 2022. The asymptomatic and progressive nature of CAD presents challenges for early diagnosis and timely intervention. Traditional diagnostic methods such angiography and stress tests are known to be resource-intensive and prone to human error. This calls for a need for automated and time-effective detection methods. In this paper, this paper introduces a novel approach to the diagnosis of CAD based on a Convolutional Neural Network (CNN) with a temporal attention mechanism. The model will be developed on an architecture that will automatically extract and emphasize critical features from sequential medical imaging data from coronary angiograms, allowing subtle signs of CAD to be easily spotted, which could not have been detected by convention. The temporal attention mechanism strengthens the ability of a model to focus on relevant temporal patterns, thus improving sensitivity and robustness in detecting CAD for various stages of the disease. Experimental validation on a large and diverse dataset demonstrates the efficacy of the proposed method, with significant improvements in both detection accuracy and processing time compared to traditional CNN architectures. The results of this study propose a scalable solution system for the diagnosis of CAD. This proposed system can be integrated into clinical workflows to assist healthcare professionals. Ultimately, this research contributes to the field of AI-driven healthcare solutions and has the potential to reduce the global burden of CAD through early automated detection.

Causal discovery algorithms are often leveraged for inferring causal relationships and recovering a causal model from data. However, causal discovery from data alone is limited by the structural constraints of the used dataset, the lack of causal logic, and the lack of external knowledge. Thus, data-driven causal discovery can only suggest possible causal relationships at best. To overcome these limitations, Large Language Models (LLMs) and knowledge systems, such as Retrieval-Augmented Generation (RAG), have been proposed as alternatives to data-driven causal discovery and as a method to augment causal discovery algorithms. Using an expert-defined causal graph of chronic lower back pain, we further propose knowledge graph based RAG systems, such as GraphRAG, as an improvement over RAG systems for augmenting causal discovery (F1 0.745), benchmarking its performance against augmenting causal discovery with an LLM (F1 0.636), augmenting causal discovery with RAG (F1 0.714), and causal discovery alone (F1 0.396). We also explore the impact of different prompting methods for causality, such as querying for the plausibility of causal relationships, the presence of statistical associations, and the existence of temporal causal relationships, as inspired by the methodology of the domain experts constructing our ground truth. Lastly, we discuss how applications of LLMs, RAG, and graph-based RAG systems can impact and accelerate the causal modeling of chronic lower back pain by bridging the gap between domain knowledge and data driven approaches to causal modeling. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=93 SRC="FIGDIR/small/26346255v1_ufig1.gif" ALT="Figure 1"> View larger version (31K): org.highwire.dtl.DTLVardef@f3387org.highwire.dtl.DTLVardef@2dforg.highwire.dtl.DTLVardef@bc839aorg.highwire.dtl.DTLVardef@63f6ea_HPS_FORMAT_FIGEXP M_FIG C_FIG

This research presents multimodal deep learning for structural heart disease prediction. We evaluated multiple deep learning architectures, including TCN, Simple CNN, ResNet1d18, Light transformer and Hybrid model. The models were examined across the three seeds to ensure robustness, and bootstrap confidence interval is used to measure performance differences. TCN consistently outperforms other competing architectures, achieving statistically significant improvements with stable performance across runs. Similarly in predictive analysis, TCN has efficient computation and stable training compared to all competing architectures. Our results show that TCN emphasizes fairness evaluation when developing deep learning models for healthcare applications.

Handheld ultrasound devices have revolutionized point-of-care diagnostics, but their effectiveness remains limited by operator dependency and the need for specialized training. This paper presents an intelligent guidance and diagnostic assistance system for the handheld wireless ultrasound device, enabling automated carotid artery and thyroid examinations through handheld operation. Drawing inspiration from the Actor-Critic framework, we implement a simulation-based reinforcement learning approach for real-time probe navigation toward standard anatomical views. The system integrates YOLOv8n-based detection networks for carotid plaque and thyroid nodule identification, achieving real-time inference at 30 frames per second. Furthermore, we propose a hybrid measurement approach combining UNet segmentation with the Snake algorithm for precise biometric quantification, including carotid intima-media thickness (IMT), lumen diameter, and lesion dimensions. Experimental validation on clinical datasets demonstrates that the proposed system achieves 91.2% accuracy in standard plane acquisition, 87.5% mean average precision (mAP) for plaque detection, and 89.3% mAP for nodule identification. Measurement results show excellent agreement with expert sonographers, with IMT measurements exhibiting a mean absolute difference of 0.08 mm. These findings demonstrate the feasibility of intelligent handheld ultrasound examination, significantly reducing operator dependency while maintaining diagnostic accuracy comparable to experienced clinicians.

BackgroundThe deployment of high-fidelity diagnostic Artificial Intelligence (AI) in resource-constrained environments is hindered by the stochastic nature of network latency and bandwidth limitations. Traditional tele-radiology relies on static cloud offloading, which introduces unacceptable latency for critical care scenarios. Zero-Touch Network and Service Management (ZSM) offers a paradigm for automated network orchestration, yet current frameworks lack application-layer awareness regarding clinical urgency and image complexity. MethodologyThis study proposes a novel Cross-Modal Latent Transformer (CMLT) integrated within a Zero-Touch Network Orchestration architecture. The system utilizes a lightweight Edge-Gating mechanism to dynamically partition inference tasks between edge nodes and cloud resources based on feature entropy. The model was trained and validated on the MIMIC-CXR (v2.0.0) (n = 377, 110) and CheXpert (n = 224, 316) datasets, employing a 70/10/20 split. ResultsThe proposed orchestration framework achieved an AUC-ROC of 0.962 [95% CI: 0.941-0.983] for Atelectasis detection, comparable to full-cloud inference, while reducing network bandwidth consumption by 64.3%. McNemars test indicated no statistically significant difference in diagnostic accuracy between the orchestrated hybrid approach and the full-precision cloud baseline (p > 0.05), despite a 120 ms reduction in mean inference latency. Clinical SignificanceBy embedding clinical feature extraction directly into the network orchestration logic, this framework enables real-time, zero-touch provisioning of diagnostic resources, facilitating reliable AI deployment in rural and bandwidth-limited clinical settings.

INTRODUCTIONConnected speech analyses can help characterize linguistic impairments in primary progressive aphasia (PPA) and classify variants, however, manual transcription of speech samples is time-consuming and expensive. Automated speech recognition (ASR) may be efficacious for transcribing PPA speech. METHODSTranscripts of picture descriptions (109 PPA, 32 healthy controls (HC)) were generated using a manual, automated (Whisper) or semi-automated approach including a quality control (QC) step. We evaluated transcript accuracy, the reliability of ASR-derived linguistic features, and classification performance. RESULTSWhisper demonstrated lowest error rates for HC, followed by semantic, logopenic and non-fluent PPA variants. Errors correlated with overall disease severity for semantic and logopenic variants. QC of Whisper outputs reduced errors and improved the reliability of linguistic features. Overall, ASR-derived features achieved better classification performance than manual transcription features. DISCUSSIONResults support the use of off-the-shelf ASR for scalable, cost-efficient transcription of PPA speech and classification.

This study presents RED-RHD, a machine learning methodology for early detection and classification of Rheumatic Heart Disease (RHD) using heart sound recordings. By leveraging OpenL3 deep acoustic embeddings, cloud-based workflows, and an ensemble of SVM and XGBoost classifiers, RED-RHD achieves an average precision of 95.62% for murmur detection (Normal vs. Abnormal) and 99.00% precision for systolic vs. diastolic murmur classification, demonstrating marked improvements over prior methods with poor cross-dataset generalization (e.g., specificity as low as 4.3% in ResNet-based approaches). These results confirm the systems robustness across diverse, noisy clinical datasets. Additionally, we introduce a novel dynamic adaptive model selection mechanism that enables the framework to automatically select the most appropriate pretrained machine learning model based on extracted heart sound features, optimizing prediction accuracy for different regional or demographic populations. By incorporating this adaptive intelligence, RED-RHD addresses population variability and supports precision diagnostics in globally diverse patient groups, advancing the potential for scalable, AI-driven auscultation in low-resource environments.

ObjectiveThis study investigates whether incorporating physiological coupling concepts into neural network design can support stable and interpretable feature learning for histopathological image classification under limited data conditions. MethodsA physiologically inspired architecture, termed CardioPulmoNet, is introduced to model interacting feature streams analogous to pulmonary ventilation and cardiac perfusion. Local and global tissue features are integrated through bidirectional multi-head attention, while a homeostatic regularization term encourages balanced information exchange between streams. The model was evaluated on three histopathological datasets involving oral squamous cell carcinoma, oral submucous fibrosis, and heart failure. In addition to end-to-end training, learned representations were assessed using linear support vector machines to examine feature separability. ResultsCardioPulmoNet achieved performance comparable to several pretrained convolutional neural networks across the evaluated datasets. When combined with a linear classifier, improved classification performance and higher area under the receiver operating characteristic curve were observed, suggesting that the learned feature embeddings are well structured for downstream discrimination. ConclusionThese results indicate that physiologically motivated architectural constraints may contribute to stable and discriminative representation learning in computational pathology, particularly when training data are limited. The proposed framework provides a step toward integrating physiological modeling principles into medical image analysis and may support future development of transferable and interpretable learning systems for histopathological diagnosis.

Current clinical risk stratification for thoracic aortic aneurysms (TAA) relies primarily on maximum diameter, which is a poor predictor of rupture. Recent fluid-structure interaction studies have identified a dimensionless "flutter instability parameter" (N{omega} ) that accurately classifies abnormal aortic growth. However, this parameter currently serves as a static diagnostic snapshot. In this work, we propose a proof-of-concept computational framework that links flutter instability to microstructural tissue damage via a coupled system of ordinary differential equations (ODEs). We model a feedback loop where flutter-induced energy dissipation drives elastin degradation and collagen remodeling, which in turn reduces wall stiffness and amplifies the instability. To address the challenge of unobservable tissue properties, we implement a Bayesian inference engine to infer model parameters. We demonstrate feasibility on a synthetic patient cohort calibrated to published clinical growth rates and diameters. Our results show that this approach can infer hidden damage parameters and capture the qualitative bifurcation between stabilizing remodeling and runaway aneurysm expansion. While validation on real patient data remains essential, this work establishes the mathematical foundation for transforming a static physiomarker into a personalized prognostic trajectory.

BackgroundArtificial Intelligence (AI) based approaches to speech analysis have the potential to assist with objective speech error analysis in aphasia but off-the shelf tools often fail to detect speech errors due to prioritizing "fluent transcription." Speech production errors (dysfluencies) are hallmark diagnostic features of the nonfluent (nfvPPA) and logopenic (lvPPA) variants of primary progressive aphasia, yet they can be challenging to detect and characterize even by expert clinicians. This study aimed to evaluate whether the novel automated lightweight Scalable Speech Dysfluency Modeling system (SSDM-L), specifically designed to detect dysfluencies, could accurately distinguish PPA variants using voice recordings of individuals reading a brief passage. MethodParticipants included a total of 104 individuals, 40 with nfvPPA, 40 with lvPPA (matched on disease severity), and 24 healthy controls who read aloud the Grandfather Passage as part of a widely used motor speech evaluation (MSE). We automatically extracted ten speech error (dysfluency) variables using SSDM-L, including insertions, replacements, and deletions at both phoneme- and word-levels, and phoneme-level prolongations and repetitions. Group differences were assessed via ANCOVAs controlling for age, education, and disease severity (MMSE, CDR sum-of-boxes). To test clinical relevance, we performed correlation analyses with MSE ratings provided by experienced speech-language pathologists (i.e., gold standard) within the nfvPPA group. Classification performance was assessed by training random forest and XGBoost machine-learning models including 5-fold cross-validation. ResultsAll individuals read the entire passage in less than five minutes. SSDM-L detected eight of the ten predefined dysfluency features at sufficient frequency to include them in subsequent analyses. All eight features distinguished PPA from controls (p<.006). Individuals with nfvPPA made more errors than the lvPPA group on every feature (all p<.023). Each feature showed a moderate positive correlation with a global MSE apraxia/dysarthria score (r=.31-.56; p<.001-.053). Together, the eight features were able to classify nfvPPA versus lvPPA at AUC=.806 (random forest) and AUC=.776 (XGBoost). DiscussionAI-based automated speech error analysis accurately distinguished nfvPPA and lvPPA variants using a brief reading task. This quick error-sensitive scalable AI system has the potential of providing a practical tool to aid diagnosis in aphasia and motor speech disorders.

Computational growth and remodeling (G&R) models have been extentively used to investigate abdominal aortic aneurysm (AAA) progression and to support clinical decision-making. However, the development of robust predictive models is often limited by the scarcity of large-scale longitudinal imaging datasets. In this study, we propose a physics-based G&R framework to simulate AAA shape evolution and generate a virtual cohort of aneurysms, thereby addressing data limitations and enabling integration with data-driven machine learning approaches for growth prediction. The proposed arterial G&R model incorporates key mechanisms influencing aneurysm progression, including elastin degradation and stress-mediated collagen production. A modified elastin degradation formulation was introduced to generate realistic aneurysm geometries exhibiting clinically relevant features such as asymmetry and tortuosity. By systematically varying parameters governing elastin damage and collagen production, 200 distinct G&R simulations were performed to produce a diverse set of AAA geometries. The dataset was further expanded using kriging-based spatial interpolation to construct a large in silico cohort. The synthetic dataset, combined with longitudinal imaging data from 25 patients, was used to train and validate four machine learning models: Deep Belief Network (DBN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). A two-step training strategy was adopted to predict maximum aneurysm diameter and growth rate based on prior geometric characteristics. The LSTM model achieved the highest performance for maximum diameter prediction (R{superscript 2} = 0.92), while the RNN demonstrated strong overall performance (R{superscript 2} = 0.90 for maximum diameter and 0.89 for growth rate). The DBN and GRU models also showed competitive predictive capability. Overall, this study demonstrates that integrating physics-based G&R simulations with machine learning enables accurate prediction of AAA growth and maximum diameter. The proposed framework provides a scalable strategy for augmenting limited clinical datasets and offers a promising tool to support personalized risk assessment and treatment planning.

Respiratory monitoring in daily-life settings is important for health assessment, yet extracting physiologically interpretable information from breathing signals under natural conditions remains challenging, as breathing is inherently dynamic and strongly modulated by behavior. Here, a portable breathing monitoring device based on a flexible lead zirconate titanate sensor is developed to address this challenge. By exploiting polarity-opposed piezoelectric and pyroelectric responses through sensor orientation, the recorded breathing waveform exhibits a characteristic dual-component structure, consisting of a narrow transient spike followed by a broad quasi-steady peak within each breathing phase. This intrinsic waveform structure enables phase-resolved quantification of how breathing effort is distributed between transient and quasi-steady components during inhalation and exhalation. Pilot measurements in healthy subjects and patients with chronic obstructive pulmonary disease or asthma reveal systematic shifts toward transient-enhanced breathing in patients, providing clearer differentiation than conventional descriptors based on breathing duration or amplitude. By transforming complex breathing dynamics into stable and physiologically meaningful signal components under daily-life conditions, this dual-response sensing approach enables more robust access to function-related changes in natural breathing.

We present a deep learning model that predicts left atrial (LA) volume from standard 12-lead ECG recordings and basic patient data. This approach offers a low-cost, scalable alternative to MRI-based LA volume measurement, which remains the clinical gold standard but is often inaccessible. Our model performs regression directly on LA volume targets and leverages Shapley values to provide interpretable feature importance. Results highlight the predictive value of ECG signals and demonstrate that patient features such as weight and height contribute meaningfully to the estimation.

The COVID-19 pandemic has presented severe challenges in understanding and predicting the spread of infectious diseases, necessitating innovative approaches beyond traditional epidemiological models. This study introduces an advanced method for automated model discovery using the Sparse Identification of Nonlinear Dynamics (SINDy) algorithm, leveraging a dataset from the COVID-19 outbreak in Thuringia, Germany, encompassing over 400,000 patient records and vaccination data. By analysing this dataset, we develop a flexible, data-driven model that captures many aspects of the complex dynamics of the pandemics spread. Our approach incorporates external factors and interventions into the mathematical framework, leading to more accurate modelling of the pandemics behaviour. The fixed coefficient values of the differential equation as globally determined by the SINDy were not found to be accurate for locally modelling the measured data. We therefore refined our technique based on the differential equations as found by SINDy, by investigating three modifications that account for recent local data. In a first approach, we re-optimized the coefficient values using seven days of past data, without changing the globally determined differential equation. In a second approach, we allowed a temporal dependence of the coefficient values fitted using all previous data in combination with regularization. As a last method, we kept the coefficients fixed to the original values but augmented the differential equation with a small neural network, locally optimized to the data of the past week. Our findings reveal the critical role of vaccination and public health measures in the pandemics trajectory. The proposed model offers a robust tool for policymakers and health professionals to mitigate future outbreaks, providing insights into the efficacy of intervention strategies and vaccination campaigns. This study advances the understanding of COVID-19 dynamics and lays the groundwork for future research in epidemic modelling, emphasising the importance of adaptive, data-informed approaches in public health planning.

Annotating seizure onset and spread in intracranial EEG is essential for epilepsy surgical planning, yet manual annotation is unreliable and cannot scale to large datasets. We introduce Neural Dynamic Divergence (NDD), an unsupervised framework that detects seizure activity by measuring deviation from patient-specific baseline neural dynamics using autoregressive models. NDD requires no labeled training data and adapts to individual patients, channels, and brain states. Validating against expert consensus annotations from 46 seizures, NDD achieves human-level agreement ({phi} = 0.58 vs. inter-rater{phi} = 0.64) and outperforms existing algorithms on 1,019 seizures with soft labels (AUROC = 0.87). We demonstrate clinical utility by automatically annotating 2,017 seizures, revealing that seizure spread patterns distinguish epilepsy subtypes and predict surgical outcomes. NDD also generalizes to continuous ICU scalp EEG monitoring (AUROC = 0.77). We provide NDD as an open-source Python package to enable scalable seizure annotation across research centers.

Abdominal aortic aneurysms (AAA) affect more than 1% of adults over 50 and carry significant mortality risk. Current surveillance relies on intermittent imaging (ultrasound or MRI) at 6-24 month intervals, which may miss rapid growth acceleration between visits. We investigate the feasibility of continuous aneurysm diameter tracking using photoplethysmography (PPG) signals. Using a one-dimensional hemodynamic model that simulates pulse wave propagation from the heart to the digital artery, we demonstrate that while single-observation diameter estimation is fundamentally limited by noise and confounding variables, aggregating thousands of observations over one or more days may achieve sub-millimeter precision. Specifically, the lower bound error analysis shows diameter uncertainty decreases to 0.7 mm with 1,600 measurements under baseline noise conditions. We validate this approach through 12- month tracking simulations of eight virtual patients with constant and accelerating growth rates, achieving root-mean-square tracking errors of [~]0.3 mm. Furthermore, we demonstrate that patient-specific model calibration from clinical measurements, despite yielding imperfect parameter estimates, still enables accurate diameter tracking (median RMSE = 0.49 mm across 50 virtual patients). These results suggest that wearable PPG monitoring could complement traditional imaging for aneurysm surveillance, potentially enabling earlier detection of growth acceleration and more timely clinical intervention. Data and Code AvailabilityAll data produced in the present study and code for generating said data are available upon reasonable request to the authors. Institutional Review Board (IRB)This research does not require IRB approval since it is not "human subjects research" as it does not include activities that involve interaction with individuals or access to identifiable private information.

Medicine historically separates abstract clinical reasoning from physical intervention. We bridge this divide with MedOS, a general-purpose embodied world model. Mimicking human cognition via a dual-system architecture, MedOS demonstrates superior reasoning on biomedical benchmarks and autonomously executes complex clinical research. To extend this intelligence physically, the system simulates medical procedures as a physics-aware model to foresee adverse events. Generating and validating on the MedSuperVision benchmark, MedOS exhibits spatial intelligence for reasoning and action. Crucially, we demonstrate that this platform democratizes clinical expertise and narrows the performance gap between junior and senior physicians. MedOS transforms clinical intervention towards a collaborative discipline where human intuition and machine intelligence co-evolve.

Background: Pressure volume (PV) loop analysis remains the gold standard for assessing the intrinsic global diastolic properties of the left ventricle (LV). Traditional fitting techniques rely on local, phase-constrained fittings and are limited due to their sensitivity to noise, landmark selection, violation of assumptions, and non-convergence. Objective: To develop and validate DIAPINN, a physics-informed neural network (PINN) framework capable of calculating intrinsic diastolic properties of the LV from measured instantaneous PV data, combining mechanistic interpretability with machine learning flexibility. Methods: Instantaneous LV diastolic pressure was modeled as the sum of 1) time-dependent relaxation-related pressure and 2) volume-dependent recoil and stiffness-related pressures. DIAPINN was trained using time, LV pressure and volume as inputs, enforcing data fidelity, model consistency, and physiological plausibility within the loss function. Performance was evaluated in 4,000 Monte Carlo simulations of LV PVloops, and in clinical data from 59 patients who underwent catheterization (39 with heart failure and normal ejection fraction and 20 controls). DIAPINN derived indices were compared to those obtained from a previously validated global optimization method (GOM). Results: On the simulation data, DIA-PINN accurately recovered all constitutive indices (intraclass correlation coefficients near unity) and improved GOM performance. On the clinical data, diastolic indices derived using DIA-PINN strongly correlated with GOM estimates (R>0.90, p<0.001) but were insensitive to initialization. DIAPINN performed best under vena cava occlusion, as varying preload improved parameter identifiability. Conclusions: When applied to instantaneous pressure volume data, a generalizable PINN framework, DIAPINN, provides an improved method for assessing global intrinsic diastolic properties of cardiac chambers.

Restoring communication for people with dysarthria secondary to pontine stroke remains a critical challenge. Intracortical brain-computer interfaces (iBCIs) have demonstrated great potential for speech restoration in people with amyotrophic lateral sclerosis (ALS), with 1-24% word error rates (WERs) on a 125,000-word vocabulary. In pontine stroke, electrocorticography (ECoG) BCIs achieved 25.5% WERs with a smaller 1,024-word vocabulary. Whether intracortical BCI performance improvements extend to people with pontine stroke-induced dysarthria remains unclear. Here, we show that neural activity from a single 64-channel microelectrode array in orofacial motor cortex can predict attempted speech in a person with pontine stroke more accurately than prior ECoG BCI work and comparably to prior iBCI work. We trained a neural network decoder to predict phoneme probabilities from spiking rates and spike-band power as BrainGate2 participant T16 mimed (mouthed without vocalization) sentences from a large vocabulary. A series of language models converted these probabilities into word sequences. This decoding architecture has remained stable more than two years post-implantation, achieving a median 19.6% WER with a 125,000-word vocabulary and a median 10.0% WER with a 1,024-word vocabulary (a 60.8% reduction over prior ECoG studies). This framework also generalized beyond cue repetition, enabling T16 to communicate spontaneously via the iBCI in a question-and-answer setting with a 35.2% WER. These results demonstrate that brain-to-text decoding from a small patch of cortex can outperform ECoG-based systems in individuals with pontine stroke and is comparable to early speech iBCIs in individuals with ALS.

In previous work, we achieved state-of-the-art performance on ChestX-ray14 (ROC-AUC 0.940, F1 0.821) using pretraining diversity and clinical metric optimization. Applying the same methodology to CheXpert, we received similar results when using NLP valuation and test data--but when evaluated against expert radiologist labels, performance was only 0.75-0.87 ROC-AUC. The models had learned to match the automated NLP labeling system, not to diagnose disease. This paper documents our investigation into this failure and our suggested resolution. We identify the NLP-to-Expert generalization gap: a systematic divergence between models optimized on labels extracted from radiology reports and their agreement with board-certified radiologists. More surprisingly, we discovered that directly optimizing for small expert-labeled validation sets can be counterproductive-- models with lower validation scores often generalized better to held-out expert test data. Four findings emerged: First, expert-labeled images for at least the validation and testing datasets, even if not for training, were vital in revealing the gap between NLP agreement and diagnostic accuracy. Without them, our models appeared excellent while failing to generalize to clinical judgment. Second, less training is better. Short training (1-5 epochs) outperformed extended training (60+ epochs) because longer training doesnt improve the model--it memorizes the labelers mistakes. Third, ImageNet features are sufficient. Freezing the pretrained backbone and training only the classifier achieved 0.891 ROC-AUC--matching models with full fine-tuning. The rapid convergence we observed wasnt the model learning chest X-ray features; it was the classifier calibrating to already-sufficient visual representations. Fourth, regularization beats optimization. Label smoothing and frozen backbones--methods that prevent overfitting--outperformed direct metric optimization on small validation sets. The 200 expert-labeled validation images in CheXpert are too few to optimize directly; they are better used as a compass than a target. With these insights, we improved from 0.823 to 0.917 ROC-AUC, exceeding Stanfords official baseline (0.907).