Research Watch · Baseline selection

Five studies. Five different claims.

The first Spinal Cord AI Research Watch separates phenotype discovery, prediction, segmentation, detection, and evidence synthesis before asking whether any result is clinically usable.

Selection date: September 19, 2026
Entries: Five source-checked studies
Review boundary: Specialist review not claimed

1 of 5

entries tested performance at an independent center.

2 abstract-only

assessments are labeled so inaccessible details are not guessed.

0 clinical utility trials

showed that using the model improved care or patient outcomes.

01 · Chiari I & syringomyelia · Phenotype discovery

Can AI identify reproducible presenting phenotypes?

Three clusters reappeared in a cohort separated by enrollment site—but they describe presentation, not prognosis or treatment response.

Publication: Peer-reviewed journal article
Study design: Multicenter registry cohort
Validation: Site-separated replication

Researchers analyzed 1,340 people with Chiari I and syringomyelia from the Park-Reeves registry. They reduced 582 candidate variables to 33 using data-driven selection plus clinician input, then applied K-modes clustering. Internal and external cohorts were separated in advance by enrollment site, and the clustering process produced three similar patterns in the external cohort: headache-predominant, younger/comorbidity-heavy, and larger-syrinx/spinal-deformity presentations. This is stronger than discovering clusters and retesting them in a random split. It still does not show that the phenotypes have different natural histories or respond differently to decompression. Exact center count, split sizes, uncertainty, and detailed replication statistics were not available in the accessed abstract.

Patients: 1,340
Images / scans: Not reported in accessed abstract
Endpoint: Presenting phenotype
Clinical-use evidence: None reported

Gupta et al., Neurosurgery · DOI 10.1227/neu.0000000000003249 · First online November 18, 2024; print 2025 · Abstract and publisher metadata checked September 19, 2026.

02 · Chiari I & syringomyelia · Outcome prediction

Can preoperative data predict delayed syrinx shunting?

A large EHR cohort produced an AUC of 0.87, but the accessible report does not establish geographic, site-held-out, or temporal external validation.

Publication: Peer-reviewed journal article
Study design: Retrospective multicenter EHR cohort
Validation: Held-out set; external status unclear

The TriNetX analysis included 3,112 patients who underwent posterior fossa decompression for Chiari I with syringomyelia during 2010–2020 and had at least two years of follow-up. Unplanned syrinx shunting occurred in 271 patients (8.7%). Among five classifiers using 30 preoperative variables and synthetic oversampling, CatBoost reached an AUC of 0.87, sensitivity of 0.78, and specificity of 0.82 in what the abstract calls an independent validation set. The accessible text does not show whether that set was separated by site, geography, or time, and calibration was not reported. EHR coding, selection, and oversampling require full-text scrutiny. Feature associations—including surgical delay—must not be interpreted as proof that changing timing will prevent shunting.

Patients / events: 3,112 / 271 shunts
Centers: Not reported in accessed abstract
Time horizon: Delayed shunt; median follow-up 5.1 years
Calibration: Not reported in accessed abstract

Abo Kasem et al., Journal of Neurosurgery: Spine · DOI 10.3171/2026.6.SPINE26351 · Published September 4, 2026 · Abstract checked September 19, 2026.

03 · Intramedullary tumors · Imaging analysis

Can deep learning separate tumor, edema, and cavity?

The open pipeline is useful research infrastructure, but one-center random splits do not establish transportability or clinical benefit.

Publication: Peer-reviewed journal article
Study design: Retrospective single-center imaging study
Validation: Repeated internal random splits

A cascaded U-Net processed paired preoperative contrast-enhanced T1- and T2-weighted MRI from 343 patients at Beijing Tiantan Hospital: 101 astrocytomas, 122 ependymomas, and 120 hemangioblastomas. Across 12 random dataset splits, combined tumor, edema, and cavity segmentation reached a Dice score of 76.7% ± 1.5%; tumor-only segmentation was lower at 61.8% ± 4.0%. True-positive detection exceeded 87% for each tissue class. The patient count and scan count were both 343, and the implementation was released through Spinal Cord Toolbox. Its openness is valuable, but all data came from one institution and tissue boundaries were often ambiguous. No independent-center test or prospective workflow comparison showed that the segmentation changes diagnosis, surgical planning, monitoring, or patient outcomes.

Patients / scans: 343 / 343 paired MRI studies
Centers: One
Reference: Quality-controlled manual segmentation
External validation: None

Lemay et al., NeuroImage: Clinical · DOI 10.1016/j.nicl.2021.102766 · Published 2021 · Full text checked September 19, 2026.

04 · Degenerative cervical myelopathy · Imaging analysis

Does cord-compression detection travel to another center?

Yes, for agreement with expert MRI annotations in a 94-patient external test—but not yet for diagnosis, management, or outcomes.

Publication: Peer-reviewed journal article
Study design: Retrospective two-center imaging study
Validation: Independent-center external test

The YOLO11 study used 1,431 sagittal T2-weighted images from 735 patients evaluated for suspected degenerative cervical myelopathy. Development and internal testing used 641 patients at one center; the external set contained 94 patients and 175 images from a second center. Against expert annotations, external mAP50 was 0.944 (95% CI 0.934–0.953), with reported class accuracy of 91% for normal and 92% for compression. This is the clearest transportability test in the baseline selection, and patients—not images—were kept separate across cohorts. The reference standard was still physician labeling of the same MRI task. The study did not test clinical symptoms, referral decisions, treatment, prospective workflow performance, or patient outcomes, and two centers cannot represent every scanner and case mix.

Patients / images: 735 / 1,431
External test: 94 patients / 175 images
Endpoint: MRI cord-compression detection
Patient-outcome evidence: None

Du et al., Spine · DOI 10.1097/BRS.0000000000005639 · Published May 1, 2026 · Full text checked September 19, 2026.

05 · Spinal cord injury · Evidence synthesis

How ready are SCI functional-outcome models?

Across 19 cohort studies, every study was rated high risk of bias and only one included external validation.

Publication: Peer-reviewed systematic review
Study design: TRIPOD and PROBAST appraisal
Validation finding: 1 of 19 studies external

This systematic review evaluated 19 cohort studies and 42 machine-learning models for neurological recovery, walking, independence, activities of daily living, and other functional outcomes after spinal cord injury. All 19 studies were judged high overall risk of bias. Sixteen used internal validation, two reported development only, and just one included external validation. No study justified sample size, only six described imputation or another missing-data method, and calibration was rarely reported. Because outcomes, horizons, predictors, and cohorts differed, the authors did not pool a single performance estimate; a unique patient total should not be manufactured from potentially overlapping evidence. The search ended May 14, 2025, so later studies are outside its scope. The review supports research potential—not routine clinical implementation.

Studies / models: 19 / 42
Unique patients: Not pooled
Risk of bias: High in 19 of 19 studies
Search end date: May 14, 2025

Liu et al., JMIR Medical Informatics · DOI 10.2196/84980 · Published June 23, 2026 · Full text checked September 19, 2026.

Selection method

A dated baseline, not a league table.

Search-assisted discovery covered Chiari I and syringomyelia, intramedullary tumors, degenerative myelopathy, spinal cord injury, and spinal vascular disease. Forty candidate web sources across four workstreams were reviewed alongside bounded PubMed searches, then deduplicated by DOI and PMID. The vascular search produced no study that met the public inclusion threshold.

Entries were selected for clinical relevance and methodological contrast, not because their metrics are directly comparable. Publication status, study design, and validation are shown separately; patient and image counts are not merged. Two entries rely on abstracts because full text was not accessible. No composite evidence score is assigned.

AI-assisted search, extraction, and drafting; editorial responsibility: Neurosurgery.ai under anonymous site-name attribution. Specialist clinical review is not claimed. Corrections and publication changes will be reconciled with the same record rather than silently replacing it. Editorial policy · Updates & corrections.