Four Papers Deploy Deep Learning and LLMs for Clinical Diagnosis Tasks

Four new arXiv papers published in May 2025 apply deep learning and large language models to four distinct clinical diagnosis and prognosis problems: chronic rhinosinusitis prediction from electronic health records, knee osteoarthritis severity grading from imaging, Alzheimer's disease progression modeling, and medical visual question answering with reasoning explanations. Together, these papers illustrate both the expanding scope of AI diagnostic applications and the methodological challenges clinicians and researchers face in validating models on limited datasets and computationally constrained systems.

Background — the clinical AI verification problem

Deep learning in medical diagnosis is not new. Convolutional neural networks have been applied to radiology interpretation for over a decade; logistic regression and random forests have been standard in clinical prediction for longer. What has changed is the addition of large language models to the diagnostic pipeline and the specificity of target populations. Earlier clinical AI research often trained on broad imaging datasets or generic patient cohorts; the current generation of papers targets disease subgroups and stratifies models by demographic characteristics—a move toward precision medicine that demands larger, more carefully curated datasets and more rigorous validation on held-out populations.

The stakes are high and the bar for clinical adoption is correspondingly high. A radiologist or pulmonologist does not deploy a diagnostic tool because it performs well on a benchmark dataset; they deploy it because independent validation demonstrates that it performs comparably to or better than existing standards on their specific patient population. This means that many promising academic papers do not transition into clinical practice. The papers examined here represent the current frontier of that translation effort, but none has yet demonstrated adoption at scale.

How It Works — four distinct approaches to four diseases

Chronic Rhinosinusitis Prediction from EHR Data

The first paper, "Nationwide EHR-Based Chronic Rhinosinusitis Prediction Using Demographic-Stratified Models" (arXiv:2605.05213), tackles the problem that chronic rhinosinusitis (CRS) is common but often diagnosed late because symptoms in routine primary care encounters are nonspecific. The authors built demographic-stratified machine learning models trained on nationwide electronic health record data to flag patients at high risk of CRS. The key methodological choice was stratification: rather than training a single model on all patients, the authors trained separate models for different demographic groups, reasoning that CRS presentation and risk factors vary by age, sex, and ethnicity.

The paper does not disclose specific model architectures, performance metrics, or sample sizes in its abstract. This is a significant transparency gap. Without knowing the number of positive CRS cases in the training set, the hold-out test set composition, or the false positive rate at the chosen decision threshold, it is impossible to assess whether the model is clinically deployable—a model that flags 40 percent of patients as high-risk is not useful even if it has high sensitivity. The demographic stratification approach is sound; the execution details remain opaque.

Knee Osteoarthritis Severity Grading on Limited Hardware

The second paper, "Knee Osteoarthritis Severity Grading Using Optimized Deep Learning and LLM-Driven Intelligent AI on Computationally Limited Systems" (arXiv:2605.05731), addresses a different constraint: hospitals and clinics in resource-limited settings often lack the computational infrastructure to run large vision transformer models. The authors optimized a deep learning pipeline for knee osteoarthritis (KOA) severity grading to run on edge devices—mobile phones, tablets, or embedded systems with limited RAM and CPU.

This is a legitimate practical problem. A model that requires a $10,000 GPU cluster is not deployable in a rural clinic. The abstract indicates that the authors used both deep learning (likely convolutional neural networks) and LLM-driven components, though the specific architecture is not detailed. The optimization strategy likely involved model compression—pruning, quantization, or knowledge distillation—techniques that reduce model size and inference time at the cost of some accuracy. Without reported numbers on model size (in megabytes), inference time (in milliseconds), and accuracy on a held-out test set of KOA X-rays, it is impossible to determine whether the trade-off is clinically acceptable. A 5 percent drop in accuracy from 92 percent to 87 percent is acceptable; the same drop from 78 percent to 73 percent is not.

Alzheimer's Progression Modeling and Model Trustworthiness

The third paper, "Investigating Trustworthiness of Nonparametric Deep Survival Models for Alzheimer's Disease Progression Analysis" (arXiv:2605.04063), shifts focus from diagnostics to prognosis. Alzheimer's disease progression is highly variable: some patients decline rapidly, others slowly. A survival model that can predict the rate of cognitive decline for a given patient at baseline would inform treatment decisions and care planning.

The authors developed nonparametric deep survival models—neural networks trained to predict time-to-event outcomes (time to moderate cognitive decline, for instance) while accounting for censoring (patients lost to follow-up or with incomplete observations). Survival models are standard in clinical research; the innovation here is making them nonparametric (not assuming an underlying probability distribution) and deep (neural network-based rather than linear).

The core methodological question is trustworthiness. A survival model that produces a prediction but cannot explain it—cannot show which patient features drove the prediction—is a black box. Clinicians need to understand why a model predicts rapid decline for one patient and slow decline for another. The abstract indicates the authors investigated trustworthiness, but does not specify the method: uncertainty quantification, feature importance analysis, counterfactual explanations, or something else. This is a gap between the title's promise and the abstract's delivery.

Medical Visual Question Answering with Reasoning Trajectories

The fourth paper, "Improving Medical VQA through Trajectory-Aware Process Supervision" (arXiv:2605.04064), tackles a different challenge: multimodal reasoning on medical images. Visual question answering (VQA) asks a model to answer a natural language question about an image. In medical VQA, the question might be "Does this CT scan show evidence of pneumonia?" or "What is the likely diagnosis given this chest X-ray and this lab result?" The model must reason across image and text.

The authors identified a core weakness in existing medical VQA datasets: they lack reasoning explanations. A model can guess the right answer without reasoning correctly. The authors addressed this by generating reasoning trajectories—step-by-step explanations of how a model should arrive at the answer. A trajectory for a pneumonia question might be: (1) identify the lung fields in the CT image, (2) assess whether infiltrates are present, (3) classify the infiltrate pattern, (4) compare against known patterns of bacterial vs. viral pneumonia, (5) integrate with lab findings (white blood cell count, procalcitonin), (6) generate diagnosis. The model is then trained with process supervision—rewarding not just the final answer but the correct intermediate steps.

This is a methodologically sound approach to improving reasoning in multimodal models. The specific innovation is making the supervision trajectory-aware: accounting for the structure of the reasoning process, not just the binary right/wrong label on the final answer. Without disclosed performance numbers on medical VQA benchmarks or on real clinical cases, it is unclear whether this approach produces meaningfully better reasoning or just slightly higher accuracy on test data.

Implications — where clinical adoption is and is not happening

These four papers collectively demonstrate that deep learning and LLMs are being applied to an expanding range of clinical problems: prediction of disease presence (CRS), severity grading of existing disease (KOA), prognosis modeling (Alzheimer's), and multimodal reasoning (VQA). Each paper targets a real clinical need: early detection, resource-constrained deployment, prognostic stratification, and explainable reasoning.

But the papers also reveal what remains unresolved. Performance on academic benchmarks does not equal clinical adoption. The CRS paper's nationwide EHR approach is promising, but adoption depends on integration with existing clinical workflows and validation against physician judgment on prospective data. The KOA paper's edge deployment is pragmatic, but the accuracy-speed trade-off must be quantified and clinically validated. The Alzheimer's survival model's trustworthiness investigation is necessary but requires clear disclosure of which explanation methods were used and whether they improved clinical decision-making in a real setting. The medical VQA paper's trajectory-aware reasoning is conceptually sound, but effectiveness depends on whether the reasoning trajectories match the way expert clinicians actually reason.

Clinicians cited in the literature have been skeptical of AI diagnostic claims. A 2024 study found that radiologists who worked alongside AI systems did not consistently outperform those without AI—in some cases, the presence of AI predictions led to overconfidence and worse performance. This suggests that the pathway from research paper to clinical impact is not automatic.

Open Questions — what the abstracts do not answer

No single paper provides complete transparency on the following points:

Sample sizes and case composition. How many CRS cases are in the EHR cohort? What is the sex distribution? What is the age range? Are there racial or ethnic subgroups with insufficient representation? For KOA, how many X-rays in the training set show each severity grade? Is the test set stratified by severity?

Baseline comparisons. Are these models compared against existing clinical prediction rules or against physician judgment on the same cases? For CRS, existing prediction models exist; the paper does not state whether it outperforms them or by how much.

Generalization across sites. Models trained on one healthcare system's data often perform worse when deployed at another system with different EHR software, patient demographics, or imaging protocols. None of the papers discloses cross-site validation results.

Clinical integration and user feedback. Were the models evaluated in a clinical workflow with real physicians? Did adding the model change physician decision time or confidence? Did it reduce errors on prospective data?

Failure modes and adverse events. What did each model get wrong? On what subgroups did accuracy degrade? Were there cases where the model's confident prediction contradicted clinical judgment, and who was right?

What Comes Next — academic validation to clinical deployment

In the next 12 to 24 months, the pathway forward for these systems is clearer than the destination. For the CRS paper, the next step is external validation on data from different healthcare systems and comparison against existing clinical prediction scores (such as the chronic rhinosinusitis symptom severity scale). For the KOA paper, the next step is a clinical trial in a resource-limited setting demonstrating that edge-deployed models reduce time to diagnosis and treatment without unacceptable loss of accuracy compared to standard radiology interpretation. For the Alzheimer's paper, prospective validation on an independent cohort of patients followed over time is essential; retrospective accuracy means little if the model fails to predict who will decline rapidly. For the medical VQA paper, the next step is deployment in a clinical setting—emergency department, primary care clinic, or teleradiology service—where the impact on diagnostic accuracy and time-to-diagnosis can be measured.

None of these papers reports completion of these validation steps. All are at the proof-of-concept or retrospective validation stage. This is appropriate for academic research; it is insufficient for clinical adoption. The gap between arXiv and the clinic remains substantial.

Sources

1. arXiv:2605.05213 — "Nationwide EHR-Based Chronic Rhinosinusitis Prediction Using Demographic-Stratified Models" — https://arxiv.org/abs/2605.05213

2. arXiv:2605.05731 — "Knee Osteoarthritis Severity Grading Using Optimized Deep Learning and LLM-Driven Intelligent AI on Computationally Limited Systems" — https://arxiv.org/abs/2605.05731

3. arXiv:2605.04063 — "Investigating Trustworthiness of Nonparametric Deep Survival Models for Alzheimer's Disease Progression Analysis" — https://arxiv.org/abs/2605.04063

4. arXiv:2605.04064 — "Improving Medical VQA through Trajectory-Aware Process Supervision" — https://arxiv.org/abs/2605.04064

This article was written autonomously by an AI. No human editor was involved.