Medicine & Health

469 items

Neurourology and Urodynamics Sep 14, 2026
This study developed and validated the Urodynamic Report Writer App (URapp), a GPT-4o-based virtual assistant modeled on scientific literature and international urodynamic guidelines, fine-tuned under supervision by two expert urodynamicists using examinations from multiple referral centers, and evaluated by functional urology specialists on a 0-5 Likert scale, scoring high on interpretative accuracy (4.5 ± 0.2), explanation of urodynamic concepts (5.0 ± 0), adherence to evidence-based standards (4.8 ± 0.3), clarity and practicality of recommendations (4.2 ± 0.3), clinical relevance (3.8 ± 0.3), and diagnostic usefulness (3.5 ± 0.5), with all 100 consecutive urodynamic studies (median age 61 yr) receiving overall scores ≥4.
This study developed and validated the Urodynamic Report Writer App (URapp), a GPT-4o-based virtual assistant modeled on scientific literature and international urodynamic guidelines, fine-tuned under supervision by two expert urodynamicists using examinations from multiple referral centers, and evaluated by functional urology specialists on a 0-5 Likert scale, scoring high on interpretative accuracy (4.5 ± 0.2), explanation of urodynamic concepts (5.0 ± 0), adherence to evidence-based standards (4.8 ± 0.3), clarity and practicality of recommendations (4.2 ± 0.3), clinical relevance (3.8 ± 0.3), and diagnostic usefulness (3.5 ± 0.5), with all 100 consecutive urodynamic studies (median age 61 yr) receiving overall scores ≥4.
Urodynamic Report Writer (URapp): A GPT-4o-Based App for Urodynamic Interpretation and Draft Report Generation
This study developed and validated the Urodynamic Report Writer App (URapp), a GPT-4o-based virtual assistant modeled on scientific literature and international urodynamic guidelines, fine-tuned under supervision by two expert urodynamicists using examinations from multiple referral centers, and evaluated by functional urology specialists on a 0-5 Likert scale, scoring high on interpretative accuracy (4.5 ± 0.2), explanation of urodynamic concepts (5.0 ± 0), adherence to evidence-based standards (4.8 ± 0.3), clarity and practicality of recommendations (4.2 ± 0.3), clinical relevance (3.8 ± 0.3), and diagnostic usefulness (3.5 ± 0.5), with all 100 consecutive urodynamic studies (median age 61 yr) receiving overall scores ≥4.
This study developed and validated the Urodynamic Report Writer App (URapp), a GPT-4o-based virtual assistant modeled on scientific literature and international urodynamic guidelines, fine-tuned under supervision by two expert urodynamicists using examinations from multiple referral centers, and evaluated by functional urology specialists on a 0-5 Likert scale, scoring high on interpretative accuracy (4.5 ± 0.2), explanation of urodynamic concepts (5.0 ± 0), adherence to evidence-based standards (4.8 ± 0.3), clarity and practicality of recommendations (4.2 ± 0.3), clinical relevance (3.8 ± 0.3), and diagnostic usefulness (3.5 ± 0.5), with all 100 consecutive urodynamic studies (median age 61 yr) receiving overall scores ≥4.
medRxiv Sep 14, 2026
Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while Qwen led normalized medication recovery (0.381 versus 0.311; difference 0.070, 95% CI 0.026–0.114) and laboratory quintuple F1 (0.892 versus 0.822), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.
Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while Qwen led normalized medication recovery (0.381 versus 0.311; difference 0.070, 95% CI 0.026–0.114) and laboratory quintuple F1 (0.892 versus 0.822), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.
Task-dependent model selection for structured extraction from multilingual non-English clinical records
Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while Qwen led normalized medication recovery (0.381 versus 0.311; difference 0.070, 95% CI 0.026–0.114) and laboratory quintuple F1 (0.892 versus 0.822), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.
Across 193,101 Russian- and Kazakh-language stroke discharge summaries, with test cohorts of 332 section cases, 149 medication cases (2,475 reference records), and 191 laboratory cases, this study compared a multilingual encoder, locally fine-tuned Qwen3-4B models, and zero-shot GPT-5.5 on entity detection versus complete-record assembly, finding section F1 of 0.919/0.926/0.932, GPT-5.5 leading drug-name detection (0.966 versus 0.940) while Qwen led normalized medication recovery (0.381 versus 0.311; difference 0.070, 95% CI 0.026–0.114) and laboratory quintuple F1 (0.892 versus 0.822), and showing that moving from curated sections to a raw-document cascade reduced medication recovery from 0.377 to 0.204 (single window) and 0.246 (all blocks) and laboratory quintuple F1 from 0.898 to 0.
Journal of Chemical Information and Modeling Sep 14, 2026
This work proposes SSE-DDI, a framework that performs selective substructure encoding in molecular line graphs with chemical bonds as the fundamental representation units, complemented by an edge-fusion graph transformer and refined SMILES-derived Morgan-fingerprint similarity profiles; on DrugBank and Twosides under transductive and inductive settings it outperforms representative baselines across multiple metrics, with ablation and visualization analyses supporting the effectiveness of selective encoding and highlighting DDI-relevant molecular substructures.
This work proposes SSE-DDI, a framework that performs selective substructure encoding in molecular line graphs with chemical bonds as the fundamental representation units, complemented by an edge-fusion graph transformer and refined SMILES-derived Morgan-fingerprint similarity profiles; on DrugBank and Twosides under transductive and inductive settings it outperforms representative baselines across multiple metrics, with ablation and visualization analyses supporting the effectiveness of selective encoding and highlighting DDI-relevant molecular substructures.
SSE-DDI: Selective Substructure Encoding with Bond-Centered Molecular Representations for Drug-Drug Interaction Prediction
This work proposes SSE-DDI, a framework that performs selective substructure encoding in molecular line graphs with chemical bonds as the fundamental representation units, complemented by an edge-fusion graph transformer and refined SMILES-derived Morgan-fingerprint similarity profiles; on DrugBank and Twosides under transductive and inductive settings it outperforms representative baselines across multiple metrics, with ablation and visualization analyses supporting the effectiveness of selective encoding and highlighting DDI-relevant molecular substructures.
This work proposes SSE-DDI, a framework that performs selective substructure encoding in molecular line graphs with chemical bonds as the fundamental representation units, complemented by an edge-fusion graph transformer and refined SMILES-derived Morgan-fingerprint similarity profiles; on DrugBank and Twosides under transductive and inductive settings it outperforms representative baselines across multiple metrics, with ablation and visualization analyses supporting the effectiveness of selective encoding and highlighting DDI-relevant molecular substructures.
JMIR Infodemiology Sep 14, 2026
Analyzing 27,434 posts from the Reddit SuicideWatch forum with a combination of manual review, large language model-assisted keyword expansion, natural language processing-based information extraction, and human validation, this study identified posts where suicidal ideation and family abuse co-occurred and described self-disclosed age and gender, abuse types, and relationships to perpetrators: among posts with self-reported age, individuals aged 18 to 24 accounted for the largest proportion (144/313, 46.0%); among posts with self-reported gender, men represented 57.3% (86/150); physical abuse was most frequent (683/975, 70.1%; 95% CI 67.18%-72.93%), followed by emotional or psychological abuse (350/975, 35.9%; 95% CI 32.89%-38.91%) and sexual abuse (296/975, 30.4%; 95% CI 27.47%-33.
Analyzing 27,434 posts from the Reddit SuicideWatch forum with a combination of manual review, large language model-assisted keyword expansion, natural language processing-based information extraction, and human validation, this study identified posts where suicidal ideation and family abuse co-occurred and described self-disclosed age and gender, abuse types, and relationships to perpetrators: among posts with self-reported age, individuals aged 18 to 24 accounted for the largest proportion (144/313, 46.0%); among posts with self-reported gender, men represented 57.3% (86/150); physical abuse was most frequent (683/975, 70.1%; 95% CI 67.18%-72.93%), followed by emotional or psychological abuse (350/975, 35.9%; 95% CI 32.89%-38.91%) and sexual abuse (296/975, 30.4%; 95% CI 27.47%-33.
Characterizing Family Abuse in Suicidal Ideation Posts on Reddit: Large Language Model-Assisted Content Analysis
Analyzing 27,434 posts from the Reddit SuicideWatch forum with a combination of manual review, large language model-assisted keyword expansion, natural language processing-based information extraction, and human validation, this study identified posts where suicidal ideation and family abuse co-occurred and described self-disclosed age and gender, abuse types, and relationships to perpetrators: among posts with self-reported age, individuals aged 18 to 24 accounted for the largest proportion (144/313, 46.0%); among posts with self-reported gender, men represented 57.3% (86/150); physical abuse was most frequent (683/975, 70.1%; 95% CI 67.18%-72.93%), followed by emotional or psychological abuse (350/975, 35.9%; 95% CI 32.89%-38.91%) and sexual abuse (296/975, 30.4%; 95% CI 27.47%-33.
Analyzing 27,434 posts from the Reddit SuicideWatch forum with a combination of manual review, large language model-assisted keyword expansion, natural language processing-based information extraction, and human validation, this study identified posts where suicidal ideation and family abuse co-occurred and described self-disclosed age and gender, abuse types, and relationships to perpetrators: among posts with self-reported age, individuals aged 18 to 24 accounted for the largest proportion (144/313, 46.0%); among posts with self-reported gender, men represented 57.3% (86/150); physical abuse was most frequent (683/975, 70.1%; 95% CI 67.18%-72.93%), followed by emotional or psychological abuse (350/975, 35.9%; 95% CI 32.89%-38.91%) and sexual abuse (296/975, 30.4%; 95% CI 27.47%-33.
Neurosurgical Review Sep 14, 2026
This systematic review and meta-analysis pooled 5 studies covering 577 patients with confirmed trigeminal neuralgia to evaluate deep learning and machine learning models for automated trigeminal nerve segmentation and neurovascular conflict detection on MRI, finding a pooled sensitivity of 71% (95% CI: 65-76%), specificity of 92% (95% CI: 88-94%), and AUC of 0.91 (95% CI: 0.88-0.93), suggesting high diagnostic accuracy for AI models in this task.
This systematic review and meta-analysis pooled 5 studies covering 577 patients with confirmed trigeminal neuralgia to evaluate deep learning and machine learning models for automated trigeminal nerve segmentation and neurovascular conflict detection on MRI, finding a pooled sensitivity of 71% (95% CI: 65-76%), specificity of 92% (95% CI: 88-94%), and AUC of 0.91 (95% CI: 0.88-0.93), suggesting high diagnostic accuracy for AI models in this task.
Artificial intelligence in trigeminal neuralgia: trigeminal nerve segmentation and neurovascular conflict detection: a systematic review and meta-analysis
This systematic review and meta-analysis pooled 5 studies covering 577 patients with confirmed trigeminal neuralgia to evaluate deep learning and machine learning models for automated trigeminal nerve segmentation and neurovascular conflict detection on MRI, finding a pooled sensitivity of 71% (95% CI: 65-76%), specificity of 92% (95% CI: 88-94%), and AUC of 0.91 (95% CI: 0.88-0.93), suggesting high diagnostic accuracy for AI models in this task.
This systematic review and meta-analysis pooled 5 studies covering 577 patients with confirmed trigeminal neuralgia to evaluate deep learning and machine learning models for automated trigeminal nerve segmentation and neurovascular conflict detection on MRI, finding a pooled sensitivity of 71% (95% CI: 65-76%), specificity of 92% (95% CI: 88-94%), and AUC of 0.91 (95% CI: 0.88-0.93), suggesting high diagnostic accuracy for AI models in this task.
Journal of Chemical Information and Modeling Sep 14, 2026
This study evaluates state-of-the-art machine-learning-based methods for nanobody structure prediction and benchmarks various HADDOCK workflows for modeling nanobody-antigen interactions across different input nanobody ensembles and information scenarios, proposing an ensemble docking pipeline that starts from nanobody structural models predicted by AlphaFold2 and ImmuneBuilder and, provided some epitope information is available, achieves higher success rates than the AlphaFold baseline on all generated models.
This study evaluates state-of-the-art machine-learning-based methods for nanobody structure prediction and benchmarks various HADDOCK workflows for modeling nanobody-antigen interactions across different input nanobody ensembles and information scenarios, proposing an ensemble docking pipeline that starts from nanobody structural models predicted by AlphaFold2 and ImmuneBuilder and, provided some epitope information is available, achieves higher success rates than the AlphaFold baseline on all generated models.
Combining AI Structure Prediction and Integrative Modeling for Nanobody-Antigen Complexes
This study evaluates state-of-the-art machine-learning-based methods for nanobody structure prediction and benchmarks various HADDOCK workflows for modeling nanobody-antigen interactions across different input nanobody ensembles and information scenarios, proposing an ensemble docking pipeline that starts from nanobody structural models predicted by AlphaFold2 and ImmuneBuilder and, provided some epitope information is available, achieves higher success rates than the AlphaFold baseline on all generated models.
This study evaluates state-of-the-art machine-learning-based methods for nanobody structure prediction and benchmarks various HADDOCK workflows for modeling nanobody-antigen interactions across different input nanobody ensembles and information scenarios, proposing an ensemble docking pipeline that starts from nanobody structural models predicted by AlphaFold2 and ImmuneBuilder and, provided some epitope information is available, achieves higher success rates than the AlphaFold baseline on all generated models.
Nature Sep 14, 2026
A Nature news report describes a proof-of-concept study published in Nature Neuroscience in which a single surgically implanted 253-electrode array covering a fairly large area of the sensorimotor cortex, combined with artificial intelligence, simultaneously decoded attempted phrases and attempted or imagined gestures in two participants with impaired speech and movement after a brainstem stroke or with amyotrophic lateral sclerosis, producing on-screen text and driving a personalized animated avatar within seconds of the user’s intent, thereby translating both verbal and non-verbal communication through one implant.
A Nature news report describes a proof-of-concept study published in Nature Neuroscience in which a single surgically implanted 253-electrode array covering a fairly large area of the sensorimotor cortex, combined with artificial intelligence, simultaneously decoded attempted phrases and attempted or imagined gestures in two participants with impaired speech and movement after a brainstem stroke or with amyotrophic lateral sclerosis, producing on-screen text and driving a personalized animated avatar within seconds of the user’s intent, thereby translating both verbal and non-verbal communication through one implant.
‘Multifunctional’ brain implant translates speech and gestures in real time
A Nature news report describes a proof-of-concept study published in Nature Neuroscience in which a single surgically implanted 253-electrode array covering a fairly large area of the sensorimotor cortex, combined with artificial intelligence, simultaneously decoded attempted phrases and attempted or imagined gestures in two participants with impaired speech and movement after a brainstem stroke or with amyotrophic lateral sclerosis, producing on-screen text and driving a personalized animated avatar within seconds of the user’s intent, thereby translating both verbal and non-verbal communication through one implant.
A Nature news report describes a proof-of-concept study published in Nature Neuroscience in which a single surgically implanted 253-electrode array covering a fairly large area of the sensorimotor cortex, combined with artificial intelligence, simultaneously decoded attempted phrases and attempted or imagined gestures in two participants with impaired speech and movement after a brainstem stroke or with amyotrophic lateral sclerosis, producing on-screen text and driving a personalized animated avatar within seconds of the user’s intent, thereby translating both verbal and non-verbal communication through one implant.
Journal of Infection and Chemotherapy Sep 14, 2026
This pilot study collected ticks between May 2023 and December 2024 from patients presenting tick bites at 10 medical institutions in Okayama, Hiroshima, and Kagawa prefectures, supplemented with wild tick images, and developed a two-stage AI pipeline of object detection and genus-level classification in which a YOLO-based model trained on 3258 annotated public images achieved mAP@0.5 of 0.954 and ResNet50 achieved mean validation accuracy of 95% on 533 tick images covering four genera, integrating the system into a prototype smartphone application with geospatial visualization to support clinical risk assessment and public health surveillance.
This pilot study collected ticks between May 2023 and December 2024 from patients presenting tick bites at 10 medical institutions in Okayama, Hiroshima, and Kagawa prefectures, supplemented with wild tick images, and developed a two-stage AI pipeline of object detection and genus-level classification in which a YOLO-based model trained on 3258 annotated public images achieved mAP@0.5 of 0.954 and ResNet50 achieved mean validation accuracy of 95% on 533 tick images covering four genera, integrating the system into a prototype smartphone application with geospatial visualization to support clinical risk assessment and public health surveillance.
Development of an AI-Based Smartphone Application for Rapid Tick Identification and Geospatial Mapping: A Pilot Study
This pilot study collected ticks between May 2023 and December 2024 from patients presenting tick bites at 10 medical institutions in Okayama, Hiroshima, and Kagawa prefectures, supplemented with wild tick images, and developed a two-stage AI pipeline of object detection and genus-level classification in which a YOLO-based model trained on 3258 annotated public images achieved mAP@0.5 of 0.954 and ResNet50 achieved mean validation accuracy of 95% on 533 tick images covering four genera, integrating the system into a prototype smartphone application with geospatial visualization to support clinical risk assessment and public health surveillance.
This pilot study collected ticks between May 2023 and December 2024 from patients presenting tick bites at 10 medical institutions in Okayama, Hiroshima, and Kagawa prefectures, supplemented with wild tick images, and developed a two-stage AI pipeline of object detection and genus-level classification in which a YOLO-based model trained on 3258 annotated public images achieved mAP@0.5 of 0.954 and ResNet50 achieved mean validation accuracy of 95% on 533 tick images covering four genera, integrating the system into a prototype smartphone application with geospatial visualization to support clinical risk assessment and public health surveillance.
JMIR Medical Informatics Sep 14, 2026
This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.
This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.
Locally Deployed Large Language Models for AI-Assisted Outpatient Prescription Review: Crossover Study
This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.
This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.
medRxiv Sep 14, 2026
Using 5,613 autopsy cases from the Mayo Clinic Brain Bank collected between 1998 and 2023, this study fine-tuned a large language model to convert narrative gross descriptions into semi-quantitative scores for 39 features (extraction accuracy 0.95 on 200 manually annotated feature-level test examples), then classified seven neuropathologic diagnostic categories with a CatBoost classifier and a second fine-tuned LLM, both including age at death, sex, and brain weight; on a held-out test set of 562 cases CatBoost reached accuracy 0.73, kappa 0.65, and macro-average AUC 0.92, while the text-based LLM reached accuracy 0.75 and kappa 0.68, with macro-average sensitivity 0.66 for both, PSP sensitivity 0.92 and 0.93, MSA sensitivity 0.87 and 0.92, but AD-LBD sensitivity only 0.21 and 0.
Using 5,613 autopsy cases from the Mayo Clinic Brain Bank collected between 1998 and 2023, this study fine-tuned a large language model to convert narrative gross descriptions into semi-quantitative scores for 39 features (extraction accuracy 0.95 on 200 manually annotated feature-level test examples), then classified seven neuropathologic diagnostic categories with a CatBoost classifier and a second fine-tuned LLM, both including age at death, sex, and brain weight; on a held-out test set of 562 cases CatBoost reached accuracy 0.73, kappa 0.65, and macro-average AUC 0.92, while the text-based LLM reached accuracy 0.75 and kappa 0.68, with macro-average sensitivity 0.66 for both, PSP sensitivity 0.92 and 0.93, MSA sensitivity 0.87 and 0.92, but AD-LBD sensitivity only 0.21 and 0.
Diagnostic Value of Large Language Model-Extracted Gross Brain Findings in Neurodegenerative Diseases
Using 5,613 autopsy cases from the Mayo Clinic Brain Bank collected between 1998 and 2023, this study fine-tuned a large language model to convert narrative gross descriptions into semi-quantitative scores for 39 features (extraction accuracy 0.95 on 200 manually annotated feature-level test examples), then classified seven neuropathologic diagnostic categories with a CatBoost classifier and a second fine-tuned LLM, both including age at death, sex, and brain weight; on a held-out test set of 562 cases CatBoost reached accuracy 0.73, kappa 0.65, and macro-average AUC 0.92, while the text-based LLM reached accuracy 0.75 and kappa 0.68, with macro-average sensitivity 0.66 for both, PSP sensitivity 0.92 and 0.93, MSA sensitivity 0.87 and 0.92, but AD-LBD sensitivity only 0.21 and 0.
Using 5,613 autopsy cases from the Mayo Clinic Brain Bank collected between 1998 and 2023, this study fine-tuned a large language model to convert narrative gross descriptions into semi-quantitative scores for 39 features (extraction accuracy 0.95 on 200 manually annotated feature-level test examples), then classified seven neuropathologic diagnostic categories with a CatBoost classifier and a second fine-tuned LLM, both including age at death, sex, and brain weight; on a held-out test set of 562 cases CatBoost reached accuracy 0.73, kappa 0.65, and macro-average AUC 0.92, while the text-based LLM reached accuracy 0.75 and kappa 0.68, with macro-average sensitivity 0.66 for both, PSP sensitivity 0.92 and 0.93, MSA sensitivity 0.87 and 0.92, but AD-LBD sensitivity only 0.21 and 0.

Page 45 · showing 10