Informatics in Medicine Unlocked

Thyroid carcinoma is one of the most prevalent endocrine malignancies worldwide, and accurate preoperative differentiation between benign and malignant thyroid nodules remains clinically challenging. Diagnostic methods that medical practitioners use at present depend on their personal judgment to evaluate both imaging results and separate clinical tests, which creates inconsistency that leads to incorrect medical evaluations. The combination of radiological imaging with clinical information systems enables healthcare providers to enhance their capacity to make reliable predictions about patient outcomes while improving their decision-making abilities. The study introduces a deep learning framework that utilizes multiple data sources by combining magnetic resonance imaging (MRI) data with clinical text to predict thyroid cancer. The system uses a Vision Transformer (ViT) to obtain advanced MRI scan features, while a domain-adapted language model processes clinical documents that contain patient medical history and symptoms and laboratory results. The cross-modal attention system enables the system to merge imaging data with textual information from different sources, which helps to identify how the two types of data are interconnected. The system uses a classification layer to classify the fused features, which allows it to determine the probability of cancerous tumors. The experimental results show that the proposed multimodal system achieves better results than the unimodal base systems because it has higher accuracy, sensitivity, specificity, and AUC values, which help medical personnel to make better preoperative decisions.

Cutaneous squamous cell carcinoma (SCC) can be time-consuming to treat with Mohs micrographic surgery (MMS) due to the need for intraoperative frozen section (FS) preparation. Two-photon fluorescence microscopy (TPFM) can generate H&E-equivalent images from fresh tissue specimens in a fraction of this time. To determine the accuracy of TPFM for the evaluation of squamous cell carcinoma in MMS margins compared to conventional FS Mohs slide preparation. TPFM was used to image 144 first stage MMS margins from patients being treated for SCC. A Mohs surgeon reviewed 44 training images and then evaluated 100 margins. After a delay, the same surgeon evaluated the corresponding FS slides. Pairs of TPFM and FS slides were reviewed by an expert dermatopathologist to form a consensus diagnosis. Agreement with consensus diagnosis as assessed by an independent dermatopathologist. 3 margins (3%) unequivocally disagreed with the consensus on TPFM and 2 margins (2%) disagreed on FS. The sensitivity and specificity of TPFM were 95.1% and 98.2%, respectively. This study demonstrates that slide-free histology can be interpreted equivalently to conventional Mohs slide processing by both MMS surgeons and dermatopathologists with minimal training.

Current standard of care imaging practices cannot reliably differentiate among certain renal tumors such as benign oncocytoma and clear cell renal cell carcinoma (RCC), and between low and high grade RCCs. Previous work has explored using deep learning, radiomics, and texture analysis to predict renal tumor subtypes and differentiate between low and high grade RCCs with mixed success. To further this work, large diverse datasets are needed to improve model performance and provide strong evaluation sets. In this work, a dataset of 831 multi-phase 3D CT exams was curated. Each exam contains up to three contrast-enhanced CT phases. Tumor outlines or bounding boxes were annotated and registered to the image volumes. The pathology results for each tumor and relevant patient metadata are also included.

Artificial intelligence (AI) based segmentation has many medical applications but limited curated datasets challenge model training; this study compares the impact of dataset annotation quality and quantity on whole liver AI segmentation performance. We obtained 3,089 abdominal computed tomography scans with whole-liver contours from MD Anderson Cancer Center (MDA) and a MICCAI challenge. A total of 249 scans were withheld for testing of which 30, MICCAI challenge data, were reserved for external validation. The remaining scans were divided into mixed-curation and highly-curated groups, randomly sampled into sub-datasets of various sizes, and used to train 3D nnU-Net segmentation models. Dice similarity coefficients (DSC), surface DSC with 2mm margins (SD 2mm), the 95th percentile of Hausdorff distance (HD95), and 2D axial slice DSC (Slice DSC) were used to evaluate model performance. The highly curated, 244-scan model (DSC=0.971, SD 2mm=0.958, HD95=2.98mm) performed insignificantly different on 3D evaluation metrics to the mixed-curation 2,840-scan model (DSC=0.971 [p>.999], SD 2mm=0.958 [p>.999], HD95=2.87mm [p>.999]). The 710-scan mixed-curation (Slice DSC=0.929) significantly outperformed the highly curated, 244-scan model (Slice DSC=0.923 [p=0.012]) on the 30 external scans. Highly curated datasets yielded equivalent performance to datasets that were a full order of magnitude larger. The benefits of larger, mixed-curation datasets are evidenced in model generalizability metrics and local improvements. In conclusion, tradeoffs between dataset quality and quantity for model training are nuanced and goal dependent.

IntroductionIdiopathic normal pressure hydrocephalus (iNPH) is a partially reversible neurological disorder in which imaging biomarkers support diagnosis and surgical decision-making. The callosal angle (CA) is one of the most robust radiological markers of iNPH and has also been associated with postoperative shunt outcome. However, several manual measurement variants exist and artificial intelligence (AI)-based tools now enable automatic CA measurement. Materials and MethodsIn total 71 patients (40 with confirmed iNPH and 31 controls) were included. Six predefined manual methods for measuring CA were applied to preoperative 3D T1-weighted MRI and evaluated for diagnostic performance and interobserver agreement. An AI-derived automatic CA (cMRI from Combinostics) was included as a seventh method and compared with the traditional manual method (perpendicular to the bicommissural plane and through the posterior commissure). Automatic measurements were additionally assessed in pre- and postoperative scans to evaluate robustness against shunt-related artifacts. ResultsAll seven CA variants significantly differentiated iNPH patients from controls (p < 0.05). The traditional method showed the highest discriminative performance (AUC = 0.986, SE = 0.012), while alternative planes demonstrated slightly lower accuracy (AUC range = 0.957-0.978). Interobserver agreement for manual measurements was good to excellent (ICC = 0.687-0.977). Automatic CA measurements showed excellent correlation with the traditional method, preoperative ICC = 0.92; postoperative ICC = 0.96. ConclusionAlthough several CA positions perform comparably, the traditional method remains marginally superior and is best supported by the literature. Automated CA measurements closely match expert manual assessment in pre- and postoperative imaging, supporting clinical implementation.

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.

BackgroundAlthough deep learning models have improved individual PET analysis, image processing and quantification tasks, end-to-end automation from raw DICOM to quantitative clinical reporting remains limited, particularly in heterogeneous real-world settings. MethodsAs a proof-of-concept, an autonomous large language model (LLM)-orchestrated multi-tool agent for end-to-end PET/CT interpretation was developed. A reasoning-based text LLM selected appropriate series from raw DICOM, coordinated registration and SUV conversion, invoked segmentation and detection tools, generated maximum-intensity projections, called a vision-enabled LLM for interpretation, and synthesized structured draft reports. The system was retrospectively evaluated in 170 patients undergoing baseline FDG PET/CT for lung cancer staging, using expert reports as reference. ResultsThe agent successfully completed the full end-to-end workflow from raw DICOM selection to structured draft report generation without human intervention in all 170 examinations. Primary tumor detection achieved 100% sensitivity. For nodal involvement, sensitivity was 84.8% and specificity was 39.4%, whereas distant metastasis detection showed 70.2% sensitivity and 65.0% specificity. Discrepancy analysis of 58 nodal and 57 metastatic mismatch cases revealed systematic false-positive findings related to reactive or physiologic uptake and false-negative findings involving small-volume or anatomically atypical metastases. ConclusionLLM-orchestrated PET/CT agents can enable workflow-level automation from raw DICOM to quantification and structured draft reporting under real-world conditions. Although primary tumor detection was highly reliable, nodal and metastatic assessment revealed systematic limitations, supporting a collaborative role with continued expert oversight in complex clinical scenarios.

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.

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.

Neurological health score (NHS), indicating the health of brain and nervous system, helps in identifying high risk individuals, and in recommending lifestyle modifications. In the present study, we developed NHS based on genetic, lifestyle and biochemical variables associated with eight neurological disorders - dementia, stroke, Parkinsons disease, amyotrophic lateral sclerosis, schizophrenia, bipolar disorder, multiple sclerosis and migraine. UK Biobank data from Caucasian individuals was used to develop the model, and the data from individuals of Indian ethnicity was used to validate the model. Logistic regression and XGBoost algorithms were used in selecting the significant variables for the disorders. NHS developed from the selected variables was found to be very significant after adjusting for age and sex (AUC:0.6, OR: 0.95). Higher NHS was associated with a lower risk of neurological disorders and better social well-being. Highest NHS group (top 25%) showed 1.3 times lower risk compared to the rest of the individuals. Results of our study help in developing a framework for quantifying the neurological health in clinical setting.

Digital breast tomosynthesis (DBT) is a powerful imaging modality that allows for improved lesion visibility, characterization, and localization compared to conventional two-dimensional digital mammography. DBT has been increasingly adopted in screening and diagnostic settings globally, particularly for women with dense breast tissue where tissue overlap presents a significant diagnostic challenge. Here we describe DBT-2026, a real world imaging dataset with 558 DBT exams from 558 patients with breast imaging reporting and data system (BI-RADS) scores of 0, 1, or 2. Each case contains one DBT examination in combination with expert annotations and free-text radiology reports that describe the radiological findings, produced in routine clinical practice. To protect patient privacy, all images and reports have been de-identified. The dataset is made freely available to researchers for non-commercial projects to facilitate and encourage research in breast cancer imaging.

Immune-related adverse events (irAEs) affect up to 40% of patients receiving immune checkpoint inhibitors, yet their identification depends on laborious and inconsistent manual chart review. Here we developed and evaluated an agentic large language model system to extract the presence, temporality, severity grade, attribution, and certainty of six irAE types from clinical notes. Retrospectively (263 notes), the system achieved macro-averaged F1 of 0.92 for detection and 0.66 for multi-class severity grading; self-consistency improved F1 by 0.14. The best-performing configuration cost approximately $0.02 per note. In prospective silent deployment over three months (884 notes), detection F1 was 0.72-0.79. In a randomized crossover study of clinical trial staff (17 participants, 316 observations), agentic assistance reduced annotation time by 40% (P < 0.001), increased complete-match accuracy (OR 1.45; 95% CI 1.01-2.09; P = 0.045), and improved inter-annotator agreement (Krippendorffs from 0.22-0.51 to 0.82-0.85). These results demonstrate that agentic AI coupled with human verification could enhance efficiency, performance, and consistency for irAE assessment.

Sepsis is a complex condition with a time-dependent evolution. Longitudinal biomarker dynamics could help us to better characterise sepsis. We hypothesised that the kinetics of biomarkers are associated with sepsis and with the intensity of organ dysfunction, and may have predictive capacity for patient survival. This single-centre, prospective, observational study included adult patients presenting to the Emergency Department (ED) with suspected infection. Patients were included in the study if they had a National Early Warning Score 2 (NEWS 2) of 3 or higher. Blood samples were obtained at baseline, 4hs and 24 hs. Linear mixed models were constructed to analyse the association between biomarker concentrations over time, sepsis diagnosis and organ dysfunction severity. Joint models were used to evaluate the predictive ability of individual biomarker kinetics during the first 24 hours for in-hospital mortality Of 214 screened patients, 173 patients were analysed, and 137 (79%) developed sepsis. Linear mixed models revealed time-dependent decreases in IL10 ({beta} -0.016, 95%CI -0.028 to -0.004), IL1RN ({beta} -0.014, 95%CI -0.024 to -0.004), and IL6 ({beta} -0.012, 95%CI -0.024 to 0.00). Sepsis was associated with higher IL1RN ({beta} 0.378, 95%CI 0.153-0.603), and TNFRSF1A ({beta} 0.40, 95%CI 0.21-0.58); only models evaluating IL6 showed significant interaction between sepsis and time ({beta} -0.14, 95%CI -0.028 to 0.00). SOFA correlated with elevated IL10 ({beta} 0.048, 95%CI 0.021-0.075), IL1RN ({beta} 0.044, 95%CI 0.017-0.071), CCL2 ({beta} 0.046, 95%CI 0.021-0.071), TNFRSF1A ({beta} 0.050, 95%CI 0.030-0.070), and PCT ({beta} 2.63, 95%CI 1.32-3.93); the interaction between SOFA score and time was significant only for IL6 ({beta} -0.003, 95%CI -0.005 to -0.001). Joint survival models (adjusted for age and highest SOFA) identified IL8 (HR 0.655, 95% CrI 0.582-0.728), TNFRSF1A (HR 0.505, 95% CrI 0.419-0.682), and PCT (HR 1.004, 95% CrI 1.001-1.008) as predictors. ConclusionSepsis diagnosis and severity of organ dysfunction may be associated with higher levels and kinetic values of inflammatory biomarkers such as IL1RN and TNFRSF1A. IL6 levels showed a significant association for the interaction of time with both sepsis diagnosis and SOFA score. TNFRSF1A, IL8 and PCT dynamics were found to be associated with survival and could be useful in developing prognosis tools.

We propose a lesion-centric phenotype learning pipeline for interpretable breast ultrasound (BUS). Predicted lesion masks are used for mask-weighted pooling of segmentation-encoder latents, producing compact embeddings that suppress background influence; a lightweight calibration step improves cross-dataset consistency. We cluster embeddings to discover latent phenotypes and relate phenotype structure to morphology descriptors (compactness, boundary sharpness). On BUSI and BUS-UCLM with external testing on BUS-BRA, lesion-centric pooling and calibration improve separability and enable strong malignancy probing (AUC 0.982), outperforming radiomics and a standard CNN baseline. A simple rule-gated generator further improves BI-RADS-style descriptor consistency on difficult cases.

BackgroundThe Most Favored Nation (MFN) policy is a mechanism that incorporates foreign prices to determine the maximum allowable net price for any branded drug within US government-funded healthcare. Two proposed rules, the Global Benchmark for Efficient Drug Pricing ("GLOBE") (90 Fed. Reg. 60,244) for Medicare Part B and the Guarding US Medicare Against Rising Drug Costs ("GUARD") (90 Fed. Reg. 60,338) for Medicare Part D, invoke the Center for Medicare and Medicaid Innovation Centers payment and service model demonstration and waiver authority, under Section 1115A of the Social Security Act (42 U.S.C. [&#167;] 1315a), to calculate the US MFN price which is the lowest average price within a basket of specified foreign countries. Unlike voluntary manufacturer agreements, GLOBE and GUARD would mandate participation from all applicable manufacturers. MethodsWe derive MFNs potential impact on Medicare pricing from a proprietary dataset provided by IQVIA which contained net prices for the top 37 oncology products by total US sales from January 1, 2019 through June 30, 2025 ranked by total US sales in the following countries: Australia, Belgium, France, Germany, Ireland, Italy, South Africa, Spain, Switzerland, the UK, and the US. For each drug, we select the lowest GDP-adjusted international price from a basket of those countries within 60% of the US GDP per capita, adjusted for purchasing power parity, and calculate the reduction in US price required to match its MFN price, and hence the corresponding reduction in revenues under MFN. A retrospective Net Present Value (NPV) analysis is then used to address the counterfactual question of whether each drug would have been developed had MFN pricing been in place at the time of its FDA approval. ResultsUnder MFN, the average reduction in US prices across our drug cohort was 67%. Eighty-four percent of the 37 cancer drugs in our cohort evidenced a negative NPV if MFN had been in place at the time of their FDA approval and the commercial market is impacted. When the analysis is restricted to MFNs impact on Medicare, the indications for these lost drugs have a total US population of 2.4 million patients. When the analysis is combined across the Medicare and commercial markets, the loss of lead indications impacts over 15 million US patients. ConclusionsMandatory MFN policies reduce the financial incentives required to develop cancer medicines; our projections show a substantial decline in new cancer drug launches and will likely lead companies to pursue indications for populations outside Medicares authority. If so, MFN will reduce the number of new therapies for the very population the Executive Orders are allegedly designed to aid: the Medicare-aged population who require effective new therapies in areas of high unmet medical need, such as late-stage cancers. This creates the perverse outcome of a policy nominally designed to help Medicare beneficiaries by instead redirecting innovation away from their most urgent therapeutic needs.

Non-alcoholic fatty liver disease (NAFLD) is a globally prevalent hepatic condition caused by the buildup of fat in the liver. It is frequently associated with metabolic comorbidities such as hypertension, cardiovascular disease (CVD), and prediabetes. However, early detection remains challenging due to the asymptomatic progression, and existing primary diagnostic methods, such as imaging or liver biopsy, are often expensive and inaccessible in rural areas. This study proposes a two-stage, interpretable machine learning pipeline for the non-invasive and cost-effective prediction of NAFLD and its key comorbidities using routine clinical parameters. The NAFLD prediction model was developed using the XGBoost algorithm, trained on a hybrid dataset that combines real patient data with rule-based synthetic data generated by simulating clinically plausible cases. Upon NAFLD-positive prediction, three separate XGB models, trained on data labelled based on thresholds, assess individual risks for hypertension, cardiovascular disease, and prediabetes. Explainability is obtained using SHAP (SHapley Additive exPlanations), which provides insight into feature relevance, while biomarker radar plots help in the visual interpretation of comorbidities. A user-friendly Streamlit interface enables real-time interaction with the tool for potential clinical application. The NAFLD model demonstrated robust performance, while the models used for predicting comorbidities achieved perfect performance, which may be a reflection of the limited dataset size used in the second stage. This work underscores the potential of AI-driven tools in NAFLD diagnosis, particularly when combined with explainable AI methods.

BackgroundMedical students represent a critical population for photoprotection education, as future physicians responsible for skin cancer prevention counseling. However, no previous studies have characterized knowledge, attitudes, and practices (KAP) regarding photoprotection among medical students in Central America or the Caribbean. ObjectiveTo assess KAP related to photoprotection and identify associated factors among medical students at a Nicaraguan university. MethodsA cross-sectional study was conducted among 133 medical students at the Universidad Iberoamericana de Ciencias y Tecnologia (UNICIT), Managua, Nicaragua. An ad hoc questionnaire assessed sociodemographic characteristics, knowledge, attitudes, and photoprotective practices. Domain-specific and global KAP scores were calculated. Bivariate analyses examined associations with sex, academic year, skin phototype, and age. ResultsParticipants were predominantly female (73.7%), with a median age of 20 years (IQR: 18-21). Although 97.0% knew what sunscreen is and 88.0% correctly identified adequate sunscreen characteristics, only 33.1% knew the minimum recommended SPF for daily use, and 21.8% understood endogenous photoprotective mechanisms. Regular sunscreen use was reported by 39.1%, while 24.8% reported never using it. Women demonstrated significantly higher scores across all domains, with moderate effect sizes for practice (d = 0.56) and global KAP (d = 0.60). No improvements were observed across academic years (p > 0.05). Age showed weak negative correlations with practice ({rho} = -0.237; p = 0.006) and global KAP ({rho} = -0.204; p = 0.018). The primary barrier to sunscreen use was forgetfulness (49.6%). ConclusionsThis first KAP study among medical students in Nicaragua reveals a substantial gap between photoprotection knowledge and practice. Current medical training appears insufficient to promote sustained protective behaviors. Findings support integrating practical, behavior-oriented photoprotection education into medical curricula and establish a regional baseline for future interventions.

There is an unmet need to identify biomarkers of active thyroid eye disease (TED). scRNAseq revealed that orbital fibroblasts from orbital decompressions in people with TED express high levels of thyroid hormone receptors, growth factor receptors, including insulin-like growth factor 1 receptor (IGF1R), and extracellular matrix proteins including SPARC (osteonectin), whereas orbital fat endothelial cells expressed thyroid peroxidase (TPO). SPARC was significantly raised in the serum of people with thyroid disease compared to healthy controls. Furthermore, those with moderate, severe and sight threatening TED had higher SPARC levels than those with thyroid disease but free of TED or mild TED. Free-triiodothyronine (FT3) levels were positively correlated with SPARC in moderate-sight threatening TED. SPARC and IGF1 were positively correlated across people with thyroid disease alone, as well as TED. Thyroid stimulating hormone (TSH) levels were negatively correlated with SPARC in moderate-sight threatening TED. When participants were followed longitudinally, SPARC decreased after the active phase of TED. At the protein level, immunohistochemistry indicated that SPARC was heterogeneously expressed by fibroblasts in both control and TED orbital fat. SPARC is a key mediator of fibrosis and deposition of extracellular matrix and the correlation of SPARC serum levels to TED status and FT3 make it a promising biomarker of active TED.

Radiogenomics enables the non-invasive characterisation of the genomic and molecular properties of tumours, with epidermal growth factor receptor (EGFR) mutations in non-small cell lung cancer (NSCLC) being one of the most investigated applications. In this study, we evaluate radiomics, contrastive learning, and convolutional deep learning approaches to predict the EGFR mutation status from chest Computed Tomography (CT) images using the TCIA Radiogenomics dataset (n=115). Our results, using 10-fold cross validation, demonstrate the capacity of imaging models to predict mutation status from CT data in a manner consistent with existing literature. Among the evaluated methods, models integrating radiomic with clinical features achieved the best performance, with an AUC of 0.790 and AUPRC of 0.517, outperforming both contrastive learning (AUC=0.787) and convolutional architectures (AUC=0.763). Beyond methodological comparisons, we discuss the challenges related to clinical translation. Specifically, we contrast radiogenomics with conventional tissue biopsies, and identify scenarios where radiogenomics might be useful, either independently or in conjunction with other existing diagnostic technologies. Together these findings evidence the potential utility of radiogenomics EGFR models and provide direct architecture comparisons on the same dataset.

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).