Deployment is the evidence gap

A 6,193-case surgical triage deployment, twelve high-risk-of-bias cervical myelopathy models, and the 2.5% trial problem show what clinical AI must prove next.

Executive assessment

Clinical AI does not finish validation when it enters a live workflow. That is when a different form of validation begins. The important questions shift from whether a model can classify a retrospective test set to whether the deployed system reaches the right patients, exposes uncertainty, fits professional work, captures representative feedback, and ultimately improves decisions or outcomes.

Three new studies describe this translational gap from different directions. A Stanford team prospectively deployed an EHR-integrated LLM triage tool across thousands of surgical cases. A systematic review of prediction models in degenerative cervical myelopathy found that every included study had high overall risk of bias. A large linkage study found registered clinical trials for only 2.5% of FDA-authorized AI medical devices—and patient-centred outcomes for 0.2%.

Deployment is not the endpoint of evidence generation. It is the point at which workflow, human judgment, feedback coverage, and patient benefit become measurable.

A real deployment: triage first, autonomy later

Evidence: Stanford preprint; prospective, unblinded workflow evaluation—not a randomized outcomes trial.

Stanford’s SCM Navigator was built to identify medically complex neurosurgery, orthopaedic, and otolaryngology patients who might benefit from perioperative hospitalist co-management. Between September 2025 and February 2026, the EHR-integrated system triaged 6,193 cases and flagged 1,582 for possible consultation. Fourteen of sixteen attending physicians participated. Read the preprint.

The design is more instructive than the use of an LLM. It did not give a single binary answer or act autonomously. A fixed prompt and predefined code path produced one of three outputs—Affirmative, Maybe, or Negative—with a brief chart-based rationale. A second model call converted the result into a structured object for Epic. Physicians saw the recommendation, retained discretion to confirm or override it, and could document disagreement.

Among cases with recorded physician determinations, the system reported sensitivity of 0.94, specificity of 0.74, positive predictive value of 0.58, and negative predictive value of 0.97. Review of false negatives attributed two of nineteen to LLM misclassification; other discrepancies reflected gaps in criteria, workflow, data, or variation in physician practice.

Those results support a narrow conclusion: an LLM can help screen a high-volume surgical queue when the task is protocolized, the output is reviewable, and clinicians preserve control. They do not establish that the system reduces complications, length of stay, cost, or clinician burden. Physician judgment—not a patient outcome—was the prospective reference standard.

“Maybe” is a workflow, not an incomplete answer

The three-state output is a useful pattern for neurosurgical AI. A binary classifier hides borderline cases inside a threshold. A visible “Maybe” category turns uncertainty into a review queue. It can direct scarce expert attention without pretending that the system has resolved the case.

But a middle category has a cost. It needs an owner, a response time, and a rule for what happens when nobody reviews it. If “Maybe” simply accumulates, the model has transferred cognitive work rather than reduced it. Deployment evaluation should therefore report not only discrimination but also review volume, time to resolution, override patterns, missed reviews, and the clinical consequences of delay.

The missing denominator: only 17% had physician feedback

The most important limitation may be the denominator. Physician feedback was recorded for 1,077 of 6,193 triaged cases—about 17%. Cases without a recorded decision were excluded from the reported performance analysis. The headline metrics therefore describe the feedback subset, not the entire deployment population.

That does not invalidate the study. It reveals a general deployment problem: labels that depend on busy clinicians are rarely missing at random. Feedback may be more likely for unusual, disputed, visible, or clinically complex cases. If the observed subset differs from the silent majority, performance can look better or worse than performance across the full workflow.

Every clinical AI deployment therefore needs a feedback-coverage plan. Passive optional feedback is useful for discovering failure modes, but it is a weak basis for estimating population performance. A defensible design adds systematic sampling of unreviewed cases, independent adjudication of a representative subset, and explicit comparison between reviewed and unreviewed populations.

Cervical myelopathy models: wide AUCs, consistently weak evidence

Evidence: peer-reviewed systematic review and PROBAST appraisal; no study tested clinical effectiveness.

A new systematic review of prediction models for degenerative cervical myelopathy included twelve studies from China, the United States, Canada, and the Czech Republic. Reported discrimination ranged from an AUC of 0.53 to 0.93, but all twelve studies were judged to have high overall risk of bias, primarily in the analysis domain. Read the review.

Only five studies reported both discrimination and calibration, and only four completed internal and external validation. None tested whether use of the model changed a clinical decision or improved an outcome. The review concluded that parsimonious clinical models using variables such as age, baseline modified Japanese Orthopaedic Association score, and symptom duration currently offer the best balance of interpretability and performance. Radiomic and machine-learning approaches have not yet shown sufficient incremental value.

This is the difference between a promising model and a usable clinical instrument. A high AUC does not show that predicted probabilities are accurate, that a threshold is clinically sensible, that the model travels across hospitals, or that acting on the output helps a patient. In a progressive condition such as cervical myelopathy, those omissions can distort counselling, timing, and expectations even when ranking appears impressive.

The 2.5% trial problem

Evidence: peer-reviewed linkage study of FDA, trial-registry, and publication records; public evidence availability is not identical to device quality.

A PLOS Digital Health study linked 1,357 FDA-authorized AI-enabled medical devices with ClinicalTrials.gov and peer-reviewed publications. The investigators identified registered trials for 34 devices (2.5%), posted results for twelve (0.9%), and linked peer-reviewed trial publications for twelve (0.9%). Only three devices (0.2%) had trials that evaluated patient-centred outcomes. Read the study.

Thirty-two of the 34 registered trials were industry-led. The authors’ larger point is not that every AI device requires a mortality trial. The appropriate endpoint depends on intended use and risk. A segmentation tool may need highly credible technical and workflow evidence; a system that changes diagnosis or treatment should face a stronger clinical standard. The linkage nevertheless exposes how often public evidence stops before patient benefit.

Regulatory authorization, technical validity, workflow utility, clinical effectiveness, and patient benefit are different claims. They should not be collapsed into the word “validated.” The Neurosurgery.ai Index should continue recording them as separate evidence layers.

A deployment evidence ladder for neurosurgery

The three studies suggest a practical sequence. Each step answers a different question, and success at one level does not imply success at the next.

1. Retrospective technical validation. Can the system reproduce a defined reference on representative historical data, with discrimination, calibration, subgroup performance, and failure analysis?
2. External validation. Does performance survive a new institution, population, data pipeline, and clinical prevalence?
3. Silent or shadow deployment. Does the complete system run reliably on live data without influencing care, and are missingness, latency, drift, and alert volume acceptable?
4. Clinician-visible controlled use. Do clinicians understand, review, override, and act on outputs appropriately? Is feedback representative?
5. Prospective decision-impact evaluation. Does the system improve decisions, workflow, equity, efficiency, or safety compared with current practice?
6. Patient-outcome evaluation. For functions that can materially change care, does use improve outcomes that matter to patients without creating offsetting harm?

Post-deployment monitoring sits across every level. A model, prompt, retrieval source, interface, or eligibility rule can change the effective intervention. Material changes require revalidation against the claims the system makes.

What to build next

1. Make the intended claim explicit. Separate technical accuracy, workflow assistance, decision impact, and patient benefit in every product page and evidence record.
2. Use an uncertainty state with an operational owner. Define who reviews “Maybe,” how quickly, what information is required, and what safe default applies if review does not occur.
3. Treat feedback coverage as a release metric. Report the fraction reviewed, compare reviewed and unreviewed cases, and independently sample the silent majority.
4. Require calibration before risk communication. A model that ranks patients well can still communicate misleading absolute risk. Calibration, confidence intervals, and threshold rationale should be visible.
5. Register prospective evaluations before deployment. Predefine the population, comparator, endpoints, subgroup analysis, failure criteria, and monitoring window.
6. Keep the model replaceable. Preserve versioned inputs, prompts, evidence, outputs, clinician actions, and outcomes so a new model can be evaluated without rebuilding the clinical system.

Bottom line

A deployable clinical AI system is not merely one that can run on live data. It is one that makes uncertainty visible, preserves human control, measures the people and cases it fails to observe, and advances from technical performance toward decisions and outcomes.

defined claim → representative validation → shadow deployment → controlled clinical use → decision impact → patient outcomes → continuous monitoring

The evidence gap is no longer only between a benchmark and a hospital. It also lies inside the deployed workflow—between an output being available, being reviewed, being acted upon, and actually helping a patient.

Editorial note: evidence labels distinguish a preprint, a peer-reviewed systematic review, and a peer-reviewed public-record linkage study. This Brief is professional information only and is not clinical advice or clinical decision support.