Thom 1995.

Methods	Prospective, randomised, unblinded trial of HBO vs. NBO. Treatment allocation by computer‐generated random numbers within sealed opaque envelopes, not sequentially numbered.
Participants	65 patients referred from local emergency departments, within 6 hours of removal from exposure. Inclusion criteria: history of acute exposure, elevated COHb, symptoms consistent with CO poisoning. Exclusion criteria: history of loss of consciousness, active ischaemia. The two groups were similar with the main clinical features being headache, nausea and lethargy and a mean COHb 20 to 25%.
Interventions	All patients in HBO arm given 100% O2 until HBO initiated. HBO begun within 6 hours of end of exposure. HBO @ 2.8 ATA for 30 minutes, then 2.0 ATA x 90 minutes. NBO 100% O2 until all symptoms resolved (mean 4.2 +/‐ 0.3 hours). After intervention, neuropsychologic baseline testing (6 tests) performed (some up to 12 hours post‐treatment).
Outcomes	Occurrence of DNS self‐reported as (1) recurrent symptoms or (2) new symptom consistent with DNS, plus deterioration in 1 or more subtest upon retesting. Outcome assessors not blind to treatment allocation. 5 patients lost to follow up (2 control, 3 HBO). 7/30 patients in control arm had sequelae consistent with DNS vs. 0/30 patients in HBO arm.
Notes
Risk of bias
Bias	Authors' judgement	Support for judgement
Adequate sequence generation?	Unclear risk	Quote: "Patients were randomly assigned to one of two treatment arms"
Allocation concealment?	Unclear risk	Not mentioned
Blinding? All outcomes	High risk	Quote: "Neither the patients nor investigators were blinded to treatment:"
Incomplete outcome data addressed? All outcomes	Low risk	Quote: "Two patients in the ambient‐pressure group and three in the HBO group were lost to follow‐up. Two patients in each treatment group refused formal neuropsychologic retesting but denied symptoms of DNS during telephone interviews conducted over the next 3 months." Comment: While the incomplete outcomes were reported, no sensitivity analysis was performed by the authors as to whether this might have influenced the findings of this trial.
Free of selective reporting?	Unclear risk	Quote: "Deteriorations in the DNS group scores occurred in three sub‐tests: Trail Making, Digit Symbol, and Block Design. The mean changes in scores were significantly different from those of a matched control group (Figure 1). No differences were found in sub‐tests on General Orientation, Digit Span recitals, or Aphasia Screening (results not shown)." Quote: "In eight patients in the ambient‐pressure group, symptoms consistent with DNS developed. Seven of these patients also had deterioration in at least one subtest category and were considered to have DNS". Comment: There was no pre‐specified primary outcome in the paper.
Free of other bias?	High risk	Intention to treat/Power/Premature stopping‐ Quotes (from 1992 abstract): "This is a preliminary report of an on‐going study". "To date, 27 patients have been entered, and 2 lost to follow‐up, in each treatment arm". "Following hyperbaric 02 treatment, no patients have suffered deteriorations (p<0.005, χ2‐ two‐tail test)." Quote from 2005 letter: "We did publish a preliminary report on our trial in 1992. At the time of that report, 27 patients had been randomised. Four patients treated with normobaric oxygen (the control group) had neurological sequelae, whereas none among those treated with HBO had sequelae". Comment: No power/sample size calculation or stopping rule was specified in the paper. There is no feasible power calculation that would have planned a study this size. [Also a P value of <0.005 simply cannot be generated from a chi square test with 4 outcomes in 27 patients (see Buckley 2005a for further exposition on this point)]. It seems likely the trial was stopped early based upon ongoing serial analysis of the available data. There was no statistical adjustment for multiple comparisons, indicating a high probability of Type I error. In any case, premature stopping of small trials "for benefit" inevitably greatly exaggerates the observed effect for statistical reasons (Bassler 2010).