BioBERT and graph attention are paired to read clinical narratives more like a chart review than a keyword search
Automatic disease classification from free-text notes has long been hampered by the way clinicians actually write. Admission notes and progress notes are dense with acronyms, shorthand, changing problem lists, and fragments of differential diagnosis, all of which make simple pattern matching brittle. MedClassNet addresses that problem by combining BioBERT contextual embeddings with a Graph Attention Network so the model can interpret both the language inside a note and the relationship between similar patients. In practice, that means the system is not limited to isolated words or phrases; it also learns how patients cluster together when their textual records carry related clinical signals.
The study used an annotated corpus of 1,748 clinical narratives from the MIMIC-III dataset, each assigned a single-label ICD-9 disease category. To keep the task clinically honest, discharge summaries, final diagnoses, explicit ICD-9 codes, and other post-diagnosis data were excluded from the input used for training. That design choice matters because it forces the model to infer disease class from the kinds of notes available earlier in care, when classification is most difficult and most clinically relevant.
The system first generates BioBERT embeddings from admission and progress notes, then constructs a similarity graph using cosine similarity between patient contextual embeddings. Patients with comparable language patterns are linked, and the graph attention layer learns which neighboring cases should carry more weight when a new record is classified. A single note can be ambiguous on its own; the graph helps resolve that ambiguity by borrowing structured context from nearby clinical neighbors.
Performance climbed when patient-level context was added to contextual language embeddings
MedClassNet achieved an accuracy of 94.2%, a precision of 94.6%, and an F1-score of 0.93 under the experimental protocol. Those numbers place the framework above baseline models in the same task setting, which supports the central premise of the work: clinical text classification improves when language understanding is paired with relational information across patients. The gain is not merely computational ornamentation. It reflects a familiar bedside reality, where two notes that look superficially different may still point to the same disease state, while two notes that share a few keywords can belong to entirely different diagnostic categories.
One sentence captures the core result plainly: the graph changed the class decision.
Because the model draws on both textual embeddings and patient similarity structure, it is better positioned to handle the instability of clinical documentation. Abbreviations vary by service line, clinicians emphasize different findings, and note templates can hide or dilute the most disease-relevant features. A transformer alone can represent the semantics of the note, but it does not automatically know which prior patients offer the best comparison. The GAT component supplies that missing relational layer.
The publication also reports SHAP-based analysis to identify clinically relevant textual features responsible for model predictions. This interpretability step is important in a domain where classification systems are often judged not only by performance, but by whether they can point to the language patterns driving their output. SHAP does not make the model clinically autonomous, but it does make the decision process more inspectable, which is essential when a system is intended to support disease categorization from chart language.
Label leakage was actively constrained to preserve a realistic prediction task
A central strength of the experimental design is the deliberate effort to avoid label leakage. By removing discharge summaries, final diagnosis statements, explicit ICD-9 codes, and other post-diagnosis information, the model is tested on data that more closely resemble the problem clinicians and informatics teams face in real time. Without that step, a classifier can appear impressive while learning from information that effectively contains the answer. Here, the input is narrowed to the earlier narrative record, where disease cues may be incomplete, indirect, or scattered across multiple notes.
That restriction raises the value of the reported performance. When a model remains strong despite the removal of obvious label-bearing text, the result suggests it is extracting signal from the clinical narrative itself rather than memorizing downstream documentation artifacts. In other words, the classification success is tied to genuine language understanding and patient-context modeling, not to a shortcut through the chart.
The disease categories themselves are single-label ICD-9 classes, so the task is narrower than a fully multimorbid real-world chart, where multiple diagnoses can overlap and evolve. Even so, the architecture is relevant beyond this benchmark. Hospitals that rely on automated coding support, surveillance pipelines, cohort assembly, or early disease triage all need systems that can work from imperfect narrative input. A framework that can classify disease classes from admissions and progress notes, while remaining sensitive to relational patterns across patients, is positioned for those settings.
No external funding supported the work, and the authors reported no competing interests. The article’s publication details place it in 2026, with the accepted version dated 13 October 2026 and publication on 3 November 2026. The broader clinical question now is not whether language models can parse medical text at all, but how well they can remain accurate when deployed on noisier records, across institutions with different documentation habits, different abbreviations, and different coding practices. MedClassNet offers a strong controlled demonstration, yet the next test is whether the same relational gains hold when the model leaves the curated confines of MIMIC-III and meets the variability of routine care.