An AUC of 0.99 Should Make a Hyperbaric Clinician Nervous

Editorial photograph of a dark clinical workstation at night: a monitor showing a scatter plot of patient outcomes divided by a curved decision boundary, with one amber outlier on the wrong side, and a brass hyperbaric chamber pressure gauge on the desk beside it, lit in teal

An area under the receiver operating characteristic curve of 0.99 is not a grade. It is a description of how cleanly two groups separated in one particular dataset. When a paper in Medical Gas Research reports that number for a model predicting which intracerebral haemorrhage patients recover after hyperbaric oxygen therapy, the honest first reaction is not admiration. It is a question about the dataset.

The paper is worth taking seriously anyway. It is one of the first machine-learning prognostic models written about a hyperbaric cohort in this journal, and it marks the moment hyperbaric medicine acquires the vocabulary other specialties have carried for a decade: nomograms, Shapley values, random forests, decision-support dashboards. That vocabulary will arrive in clinic brochures long before it arrives in a randomised trial. So it is worth knowing exactly what these models can and cannot show.

Four hundred and one patients, one perfect score

Ding and colleagues, from the First Hospital of Jiaxing in Zhejiang province, went back through the records of 401 patients with intracerebral haemorrhage who received hyperbaric oxygen at their institution. They split them by modified Rankin Scale score: 265 with a good outcome (mRS below 3) and 136 with a poor one. They then fed the clinical variables into logistic regression and a set of machine-learning classifiers, assessed performance with 10-fold cross-validation, and used SHapley Additive exPlanations to say which variables were driving the predictions.

The independent factors they identified are, to a clinician, unsurprising: Glasgow Coma Scale score, haemorrhage volume on CT, activities-of-daily-living score, treatment modality, intraventricular bleeding, and rebleeding. The best performer was a random forest, at that AUC of 0.99.

Read the list of predictors again. Every one of them is a severity marker that any experienced neurosurgeon or intensivist would use at the bedside, and one of them — treatment modality — is a decision rather than a measurement. A model that combines GCS, clot volume and functional score to predict a Rankin outcome is not discovering new biology. It is confirming that sicker brains do worse, which is useful for building a nomogram and reconciliation of expectations, and is not the same claim as “hyperbaric oxygen works for intracerebral haemorrhage.”

The design has a further constraint that the authors state plainly enough if you know where to look: all 401 patients were treated. There is no untreated comparator, no external validation cohort, and no information on how the decision to treat was made. Cross-validation resamples one hospital’s practice; it cannot tell you whether the model travels to the next hospital, let alone whether it separates the effect of the chamber from the effect of having been selected for one. An AUC of 0.99 in that setting is a signal to ask what leaked, not to book the conference talk.

The underlying treatment question is not settled either. A review of 38 studies on oxygen therapy after intracranial haemorrhage, led from Tianjin Medical University General Hospital and published in CNS Neuroscience & Therapeutics in February, found plausible mechanisms and a clinical literature that is small, heterogeneous and, in the authors’ own words, carries “controversial and potential risks,” including the possibility that hyperoxia itself harms these patients. Their conclusion asks for large-scale trials. The nomogram arrived first.

The inventory of fifty-three studies

Anyone who wants to know where this is heading should read the other paper in the same issue of Medical Gas Research. Francisco Epelde, an internist at Parc Taulí University Hospital in Sabadell, Spain, screened PubMed, Scopus, Web of Science and Google Scholar from January 2000 to March 2025 and found 53 eligible studies linking hyperbaric oxygen with artificial intelligence or its enabling technologies. At least 30 per cent of them appeared between 2022 and 2025.

He sorts them into four groups. Algorithms that optimise protocols using multicentre data. Algorithms that select patients using multi-omics and imaging biomarkers. Biosensor-driven adaptive control during treatment. And predictive safety analytics for oxygen toxicity and barotrauma. He is candid about why the field is receptive: hyperbaric outcomes are “constrained by heterogeneous protocols, limited patient selection tools, and the absence of real-time physiological guidance.” In other words, the same weakness that makes the therapy hard to evaluate makes it a promising target for software.

His priority list is the part worth pinning to a wall: hybrid clinical trials, interoperable registries, explainable and bias-aware models, and validation that lines up with regulatory expectations. Not one of those is a modelling problem. They are all data-infrastructure problems, and they are the reason a model can score 0.99 in Jiaxing and still be untransferable.

When the treatment itself is one of the predictors

Two carbon-monoxide papers make the confounding problem concrete in a way the intracerebral haemorrhage study does not.

In BMC Neurology in February, a Chongqing group led by Gong built a four-variable logistic model for delayed encephalopathy after acute carbon monoxide poisoning from 102 patients, 15 of whom developed it. The model performed well: AUC 0.933, a Brier score of 0.066 after 1,000 bootstrap resamples, and decision-curve analysis the authors present as supporting clinical usefulness. Its four variables were length of stay, comorbidities, carboxyhaemoglobin, and hyperbaric oxygen therapy after discharge.

That last variable is the interesting one. Receiving hyperbaric oxygen after discharge is not plausibly a cause of delayed encephalopathy. It is far more plausibly a marker of how severely a patient was affected, how closely they were followed, and how well resourced their care was. The model is statistically sound and inferentially treacherous: it would happily tell a clinician that post-discharge sessions raise risk.

The same pattern recurs at larger scale. Zhu and colleagues, working across Nanjing and Suqian, trained sixteen machine-learning models on 360 carbon monoxide patients, 103 of whom (28.6 per cent) developed delayed neurological sequelae. A random forest with synthetic minority oversampling reached an AUC of 0.89, with sensitivity 0.9 and specificity 0.8, and their Shapley analysis ranked the Glasgow Coma Scale, hyperbaric oxygen therapy, kidney function, immune response, liver function and clotting among the drivers. And in severe sudden sensorineural hearing loss, a Korean group built a 12-feature neural network on 231 patients that reached 89.36 per cent test accuracy and an AUC of 0.87, with “10 or more hyperbaric sessions” carrying high importance.

That last paper deserves credit for saying the quiet part. Hyperbaric oxygen was included, its authors write, “as an exposure feature rather than inferring causal treatment effects,” and they add that “these findings do not establish the therapeutic efficacy of HBOT.” They also list what is missing: external validation and better adjustment for confounding.

This is the discipline the whole genre needs. A model that includes the treatment among its predictors is measuring the prescribing pattern as much as the disease. When a clinic later advertises that its algorithm “predicts your outcome,” the honest reading is that it predicts what happened to similar patients at one centre in a particular era, under a particular selection rule.

Where the labels are objective, the algorithms get better

The strongest algorithmic work in this literature is not prognostic at all. It is surveillance, and it targets the two hazards that make hyperbaric exposure dangerous: central nervous system oxygen toxicity and inert-gas narcosis.

In Communications Medicine in 2024, a team spanning the University of Connecticut and the Navy Experimental Diving Unit studied 26 divers across 50 chamber exposures, exercising head-out in water while breathing 100 per cent oxygen at 35 feet of seawater — a partial pressure of 2.06 ATA — for up to two hours. Thirty-two exposures produced symptoms judged definitely or probably due to central nervous system oxygen toxicity; eighteen did not. Using time-varying spectral indices of electrodermal activity, the researchers found a significant rise in the index a mean of 57 seconds before symptoms appeared, with a median lead time of 10 seconds. Sensitivity was perfect at 1.0. Specificity was 0.48.

A specificity of 0.48 means roughly half of the alarms are false. As a product that would be unacceptable; as a research result it is the right kind of promising, because the label — a symptom a physician can observe — is objective and the signal precedes it. Improve the detector and you have a genuine early warning for the dose ceiling.

The counterweight comes from Auckland. James and colleagues recorded 32-channel EEG from 45 divers at rest across four randomised crossover hyperbaric studies, manipulated pressure and gas to create narcotic and non-narcotic conditions, extracted 2,720 features, and trained six classifiers. On a 9-participant test set, linear discriminant analysis reached a Matthews correlation coefficient of 0.50 and 82 per cent accuracy against a 76 per cent baseline. On a 12-participant hold-out set collected years later, the same approach fell to an MCC of 0.25 and 69 per cent accuracy against a 69 per cent baseline — a p-value of 0.77, which is to say, nothing.

Two things are worth taking from that. The first is scientific honesty; the Auckland team reported the collapse rather than burying it, and noted that their outcome was gas composition rather than measured cognitive impairment. The second is the reason to be sceptical of the Jiaxing number. A careful model built on 2,720 features degraded to chance when the population changed. An AUC of 0.99 built on 401 patients from one hospital has not yet met a second hospital.

The model that already works has no patients in it

The most useful algorithmic tool in hyperbaric medicine this decade may not involve a patient at all. It is a spreadsheet that recalibrates a ventilator.

At Hennepin County Medical Center in Minneapolis, Winn and colleagues tested a Zoll Z Vent in assist-control mode inside a Class A multiplace chamber against a Michigan test lung system. Their hypothesis was that the ventilator would deliver set tidal volumes incorrectly in volume-control mode under pressure, and that the error could be corrected by an algorithm rather than by guesswork. It could. The set tidal volume needed to achieve a target volume scaled linearly with depth, and the slope of that relationship scaled linearly with the target volume itself. A second round of regressions produced an equation predicting the set tidal volume for any given target volume, PEEP and gas mix, and the team built a spreadsheet their clinicians now use, internally validated on more than 100 patients.

This works for reasons that have nothing to do with the glamour of machine learning. The target is measurable within seconds. The physics is deterministic. A wrong prediction is caught before the breath ends. Compare that to a prognostic nomogram, where the error surfaces years later and is indistinguishable from the natural history of the disease.

There is a marketing layer above all this, and it is already here. A 2025 paper in Undersea and Hyperbaric Medicine by two authors who list two hyperbaric clinics as their affiliations reported eleven school-aged athletes treated for acute concussion within ten days of injury, at 1.5 to 2.0 ATA, stepped up as tolerated, in sessions repeated until symptoms resolved. Each patient was scored before and after treatment with an FDA-cleared, artificial-intelligence-based EEG brain biomarker that returns a 0 to 100 score, with 70 or below counted as concussed. The median score went from 18 at baseline to 85 at the last follow-up, over a median of three treatments across two days.

The number is dramatic. The design is a case series of eleven patients with no control group, no blinding, and a symptom-driven endpoint, published by the clinics that delivered the treatment. And the biomarker, whatever its clearance status, is cleared to assess a concussion — not to demonstrate that a chamber treats one. That distinction is where a great deal of consumer marketing quietly goes wrong.

Regulatory reality has not moved. Hyperbaric oxygen devices are Class II devices cleared through the 510(k) pathway under Product Code CBF, and the FDA’s August 25, 2025 letter to health care providers still stands, with its reports of fires causing serious injuries and deaths, its insistence on instruction-for-use compliance, staff training, monitoring for the full duration of treatment, and its prohibition on electrical and static-generating items in the chamber. A clinic running an “AI-guided” protocol has exactly the same fire-prevention obligations as one running a paper one.

Nothing on the UHMS fifteenth-edition indication list — read again this week, all fifteen entries from air or gas embolism through avascular necrosis — covers intracerebral haemorrhage or concussion. Medicare’s National Coverage Determination 20.29 names acute carbon monoxide intoxication, decompression illness, gas embolism, gas gangrene, crush injury, progressive necrotising infections, compromised grafts and flaps, refractory osteomyelitis, osteoradionecrosis, soft-tissue radionecrosis, cyanide poisoning, actinomycosis and diabetic lower-extremity wounds meeting three criteria. Intracerebral haemorrhage and concussion are not on that list either. An algorithm does not add an indication.

The clearest sign of where this is going is a registry entry, not a paper. NCT07855848, last updated on October 2, is titled “AI-Assisted Precision Hyperbaric Oxygen Therapy Based on Dynamic BIS-EEG Feedback for Patients With Disorders of Consciousness.” It plans 200 participants, and its sponsor is listed as an industry entity, the Boston Intelligent Medical Research Center, linked to Shenzhen United Scheme Technology. It is not yet recruiting.

That is the shape of the next few years: not algorithms proving hyperbaric oxygen works, but algorithms being built into how it is prescribed, monitored and sold. The field’s own review says what would make them trustworthy — registries, external validation, explainable and bias-aware models, regulatory-aligned evidence. Until those exist, the useful question about any hyperbaric nomogram is not how high its AUC goes. It is whether the model was ever asked to separate the chamber from the decision to use it.

Sources

← Back to blog