Abiram Sandralegar

This study conducts a clinical evaluation of a secure, locally deployed, quantized large language model (LLM) for automating modified Rankin Scale (mRS) score extraction from unstructured neurosurgical notes. We retrospectively selected 103 authentic clinical letters (2007–2025) from aneurysm patients at a tertiary neurosurgical centre. To comply with data privacy constraints, an open-source reasoning LLM (Qwen3-32B) with 4-bit quantization was deployed entirely on-premises. The LLM extracted mRS scores using a zero-shot approach with custom logits processors to enforce strict JSON formatting. Performance was compared to a reference standard (attending neurosurgeons’ consensus) and parallel scoring by medical residents and students. The LLM achieved excellent agreement with the attending consensus (QWK 0.95), matching the reliability of medical students (QWK 0.95) and residents (QWK 0.93). Exact agreement was 75%, and agreement within ±1 mRS point was 96%. Bayesian analysis strongly supported statistical equivalence between the model and human raters. The computationally optimized LLM demonstrated human-level classification reliability without task-specific fine-tuning. This approach successfully addresses key patient data privacy barriers and the formatting inconsistencies typical of open-ended generative models. Securely deploying a general-purpose, quantized LLM provides a scalable pathway for extracting functional outcomes and supports FAIR-aligned data systems.

Background/Objectives: Degenerative cervical myelopathy (DCM) is the leading cause of non-traumatic spinal cord dysfunction in adults. Both anterior and posterior surgical approaches are widely used for its treatment, yet their comparative impact on clinical outcomes and complication profiles remains clinically relevant. This study aimed to describe early and one-year postoperative outcomes between anterior and posterior surgical strategies in a consecutive single-center cohort. Methods: We conducted a retrospective single-center study of 102 consecutive adult patients surgically treated for DCM between 2013 and 2024, with complete datasets and a minimum one-year follow-up. Surgical approaches were classified as anterior, posterior, or combined. Clinical outcomes were assessed using the modified Japanese Orthopaedic Association scale (mJOA), Neck Disability Index (NDI), and Numeric Rating Scale for arm pain (NRS-arm) preoperatively, at 4–6 weeks and at one year. Complications were systematically recorded and stratified by approach. Nonparametric tests were used for all comparisons (significance threshold p < 0.05). Results: Anterior procedures were performed in 82 patients (80.4%), posterior in 18 (17.6%), and combined in 2 (2.0%). No significant between-group differences were observed in neurological, functional, or pain outcomes at either time point (ΔmJOA p = 0.268; ΔNDI p = 0.632; ΔNRS p = 0.562 at one year). Complications occurred in 13 patients (12.7%), with approach-specific profiles: anterior surgery was associated with hematoma, dysphagia, and dysphonia, posterior surgery with CSF leak, wound infection, and kyphosis. No C5 palsy was recorded. Conclusions: Both anterior and posterior surgical approaches were followed by neurological and functional improvement at one year. Given the descriptive nature of the study and the baseline differences between groups, these findings should not be read as a formal comparison of effectiveness, but they reinforce the importance of individualized, pathology-driven surgical planning in DCM.

Large Language Models (LLMs) show strong potential for extracting structured information from unstructured clinical narratives. However, their adoption in healthcare is constrained by privacy requirements that necessitate local deployment, often under limited computational resources. This study systematically investigates the trade-offs between model size and numerical precision (quantization) in locally deployed, open-source LLMs applied to a clinical information extraction task. Using seven reasoning models from the Qwen3 and DeepSeek-R1 families (0.6B–70B parameters) and six quantization levels (2, 3, 4, 5, 6, and 8 bits), we evaluated performance on the classification of the modified Rankin Scale (mRS) score from 105 neurosurgery patient notes. A linear mixed-effects analysis revealed significant effects for log-scaled model size (β=-0.84, p<0.001), quantization (β=-0.12, p<0.001), and their interaction (β=0.08, p<0.001). Larger models and higher numerical precision yielded lower extraction error, but the performance plateaued at 4–5-bit precision. Our results suggest that clinically useful information extraction is achievable with moderately sized, locally deployed LLMs, supporting their feasibility for privacy-preserving healthcare applications.

Clinical narratives contain rich, detailed information that is essential for medical research but often locked behind privacy constraints due to the presence of personally identifiable information. To enable secondary use while protecting patient confidentiality, these texts must undergo de-identification. Because even a few missed instances can lead to disclosure, systems must achieve high recall. In this work, we present CONORM-DEID, a multilingual model for clinical text de-identification, and investigate how large-scale synthetic pre-training and the availability of human-annotated data impact model robustness in both English and French. We compare models trained from scratch with models incorporating synthetic pre-training across high- and low-data settings. All models trained from scratch achieve high F1-scores (97%+), but robustness, measured as precision at 99% recall, improves only when synthetic pre-training is followed by substantial human fine-tuning. When annotated data are limited, pre-training instead reduces performance, indicating negative transfer. Our code is publicly available at https://github.com/ds4dh/multilingual_deidentification.

Traumatic pericallosal artery aneurysms (TPAAs) are a rare but potentially fatal subset of intracranial aneurysms, typically resulting from blunt or penetrating head trauma. Located along the distal anterior cerebral artery (DACA), their deep midline position and proximity to the falx render them vulnerable to shear injury, technically challenging to diagnose and treat. They often present as pseudoaneurysms with a high risk of delayed rupture. We conducted a systematic review of studies published between 1995 and April 2025 using PubMed and Google Scholar. Included studies described pericallosal or DACA traumatic aneurysms and reported clinical, radiological, and outcome data. Extracted variables included patient demographics, mechanism of injury, diagnostic modality and timing, aneurysm characteristics, treatment type, and outcomes. Statistical analysis assessed predictors of neurological prognosis using the Glasgow Outcome Scale. Fifty-eight patients from 44 studies met the inclusion criteria. Blunt trauma was the most common cause (62.1%). Aneurysms were more frequently right-sided (69%), and nearly half were diagnosed within 2 weeks postinjury. Endovascular (39.7%) and microsurgical (55.2%) treatments were both commonly employed, while conservative management was rare (3.4%). No significant outcome differences were found between treatment modalities. Poorer outcomes were significantly associated with aneurysm rupture ( p < 0.01) and older age ( p ≈ 0.02), while laterality and trauma mechanism were not. Pediatric patients demonstrated more favorable outcomes. In conclusion, TPAAs are rare but clinically urgent lesions. Early vascular imaging and timely intervention are essential. Prognosis depends more on rupture status and patient age than on treatment modality. Standardized outcome reporting is needed to guide management.

Background: The management of unruptured intracranial aneurysms (UIA) requires a balance between procedural risks and the potential benefit of rupture prevention. Objective: The aim of this study is to evaluate the clinical and anatomical results of endovascular treatment for UIAs at our institution, to benchmark our practice and potential areas for improvement, and enable transparent communication with patients by providing accurate, evidence-based information. Methods: We reviewed all patients treated for an UIA between January 2017 and July 2022. Patients were grouped according to treatment technique: simple or balloon-assisted coiling; stent-assisted coiling; or flow-diverter stent placement. Clinical outcomes included perioperative mortality, transient (3 months). Anatomical results were graded with the Raymond-Roy or the O’Kelly-Marotta scales. Retreatment rates, length of hospitalization, readmissions, irradiation time, and total radiation dose were assessed. Results: In total, 169 patients with 201 UIAs underwent 187 endovascular procedures without death, 1% permanent morbidity, 8% transient morbidity, and 4% retreatment. The treatment subgroups analysis shows an absence of permanent morbidity associated with flow-diverter stenting and stent-assisted coiling. Simple/balloon-assisted coiling suffered 2.6% permanent morbidity. Transient morbidity was observed in 9%, 9.5%, and 6.6% of flow-diverter stenting, stent-assisted coiling, and simple/balloon-assisted coiling, respectively. Conclusions: In our hospital, about one in a hundred procedures caused lasting problems, while about one in 25 aneurysms needed a second treatment. These results reflect our case selection and multidisciplinary approach. Reporting them transparently helps patients understand what outcomes to expect at our hospital.