If you’ve undergone surgery—or accompanied a family member through surgery—you’ve likely witnessed this scene: The anesthesiologist asks you to count backward from ten; before you reach seven, you lose consciousness. When you open your eyes again, the surgery is over. During that intervening time, you’re unable to ask questions or respond—you simply sleep, and how deeply you sleep is entirely determined by others. Yet the question of "how deep" can go wrong in two opposite directions.
## Too Shallow vs. Too Deep: Each Carries Its Own Cost
Too shallow anesthesia may lead to the patient’s greatest fear: accidental awareness under general anesthesia. Patients may hear sounds, feel pain, and even retain memories afterward—yet remain paralyzed by neuromuscular blocking agents. A 2008 trial published in *The New England Journal of Medicine* described this as "a serious complication that may cause long-term psychological consequences."
Too deep anesthesia isn’t a cost-free error either. Two trials published in *JAMA* open with nearly identical statements: intraoperative suppression of electroencephalographic (EEG) activity indicates excessive anesthesia, and such suppression correlates with postoperative delirium. A 2026 deep-learning study states it even more plainly in its introduction—deep anesthesia is associated with delayed recovery, postoperative cognitive dysfunction, and increased long-term mortality.
How rare is awareness, really? The UK’s Fifth National Audit Project provides a detailed accounting: across all anesthesia cases, accidental awareness occurs approximately once per 19,600 cases; when neuromuscular blocking agents are used, the rate rises to about once per 8,200 cases; for cesarean deliveries, it’s roughly once per 670 cases. Two-thirds of cases do not occur during the maintenance phase but rather during the dynamic phases of induction and emergence. Identified risk factors include female sex, young adulthood, obesity, emergency surgery, obstetric, cardiac, or thoracic surgery, use of neuromuscular blocking agents, and less-experienced anesthesiologists. Risk is highly concentrated—precisely among the populations least able to afford oversights.
## Four Major Trials—and One Unproven Number
Many operating rooms display a monitor that converts EEG signals into a number ranging from 0 to 100: 0 signifies no detectable brain electrical activity; 100 signifies full wakefulness. Does watching this number actually improve patient outcomes? This question has been tested at the highest level of clinical research—four times.
The first was the 2008 B-Unaware trial led by Avidan et al., published in *The New England Journal of Medicine*: 2,000 high-risk patients were randomly assigned to either bispectral index (BIS)-guided anesthesia or end-tidal anesthetic concentration–guided anesthesia. Results showed two confirmed cases of awareness in each group—no difference. The authors’ conclusion was unequivocal: "These findings do not support the routine clinical use of BIS monitoring."
The second was the larger 2011 BAG-RECALL trial, also by the same team: 6,041 high-risk patients were enrolled, with monitors equipped with audio alarms, plus structured education and checklists. The results were even more awkward—7 confirmed awareness cases (0.24%) occurred in the BIS group versus only 2 (0.07%) in the end-tidal concentration group; superiority testing failed (P = 0.98). When combining confirmed and possible cases, the numbers were 19 (0.66%) versus 8 (0.28%), and the confidence interval for the difference did not even include zero—the direction favored the control group, precisely opposite to what proponents of the monitor had expected.
A 2019 Cochrane review pooled data from 52 trials involving 41,331 participants and reached the same conclusion as individual trials: compared with end-tidal anesthetic concentration guidance, BIS guidance showed no evidence of reduced awareness incidence—both groups hovered around 0.1%, and the certainty of evidence was rated low.
The third and fourth tests shifted focus: even if awareness cannot be prevented, can monitoring the number at least avoid overdosing and protect older adults’ brains? The 2019 ENGAGES trial, published in *JAMA*, randomized 1,232 patients aged 60 years or older undergoing major surgery—patients in the EEG-guided group received less anesthetic (median end-tidal concentration 0.69 vs. 0.80 MAC), and experienced shorter durations of EEG suppression (median 7 vs. 13 minutes). Yet the incidence of delirium within five postoperative days was 26.0% vs. 23.0%—not lower, but slightly higher—with no statistically significant difference. In 2024, the ENGAGES-Canada trial enrolled 1,140 elderly patients undergoing cardiac surgery across four Canadian hospitals: anesthetic use was similarly reduced, and median EEG suppression duration dropped from 11.7 to 4.0 minutes; no patient reported intraoperative awareness—but delirium incidence was nearly identical: 18.15% vs. 18.10%.
Not all aggregated data are so bleak. A November 2025 update published in *International Journal of Surgery* meta-analyzed 12 randomized trials involving 4,523 patients: EEG-guided anesthesia overall reduced postoperative delirium risk by approximately one-fifth (relative risk 0.81), with greater benefit observed in patients aged 60+ and in non-cardiac surgery; it was also associated with shorter hospital stays and reduced vasopressor use, while mortality showed no difference. Yet the authors themselves highlighted marked regional heterogeneity: trials from Europe, East Asia, and Australia showed benefit, whereas North American trials did not—and notably, the two delirium trials cited above were both conducted in North America.
The common thread across all four trials is this: watching the number indeed changed how clinicians administered drugs—but it did not reliably change patients’ postoperative experiences.
## Next-Generation AI: From "How Deep Is It Now?" to "How Deep Will It Be?"
Studies published in 2025 and 2026 have taken a new direction—not merely displaying current depth, but forecasting where depth will go several minutes ahead.
A deep-learning framework called AnesNet, tested on 2,803 surgical patients, issued warnings for adverse anesthesia depth states 5, 10, and 15 minutes in advance, achieving an AUROC of 0.895 for classification. A study in *British Journal of Anaesthesia* used heart-rate signals alone: analyzing intraoperative ECGs from 3,338 patients, it discriminated "shallow anesthesia" events with an AUROC of 0.953, retaining an AUROC of 0.903 even 15 minutes in advance; a streamlined model using only 27 features accelerated computation 110-fold. Another study extracted transient alpha spindle events from frontal EEG and claimed to forecast drug overdose—signaled by isoelectric suppression—up to 90 seconds in advance, with 96% accuracy. Still another study subdivided sedation depth from a binary state (awake/asleep) into three states (awake, transition, unresponsive), achieving classification accuracies of 93.93% and 97.42% under propofol and midazolam, respectively.
There are also lower-cost approaches. One study fed raw EEG from 143 patients directly into a neural network and found performance comparable to complex signal-decomposition pipelines—while requiring less computation, which the authors argue makes it more accessible for resource-limited hospitals. A subsequent commentary countered that lighter preprocessing does not necessarily translate to shorter intrinsic latency. Another study applied unsupervised clustering to automatically categorize over 16,000 EEG data points into shallow, appropriate, and deep states—entirely without manual labeling. On the automation front, a 2026 systematic review in *British Journal of Anaesthesia*, synthesizing 17 trials involving 1,898 patients, found that closed-loop infusion systems increased time spent within the target depth range by 17.6%—equivalent to ~20 extra minutes of appropriate depth during a two-hour operation; time spent too deep (BIS < 40) significantly decreased, time spent too shallow did not increase, and extubation occurred on average 1.68 minutes earlier.
## The Circularity in Labels: Accurate Prediction ≠ Correct Prediction
Stacking these achievements reveals an easily overlooked issue: most new models are trained using BIS itself as the "ground truth"—predicting its trajectory, predicting its boundary violations, or regressing toward it. This amounts to training AI to predict ever more accurately a number whose clinical value remains unproven after four major trials. Accurate prediction does not equal correct prediction.
Fortunately, smarter studies are already breaking this cycle: some use isoelectric suppression as the endpoint; others use patients’ behavioral responses as labels; still others abandon labels altogether, letting data self-cluster. Their bet is on signals more directly linked to physiological outcomes. There remains, however, an unavoidable structural challenge: awareness events are extremely rare (~1 per 19,600 cases), and a single delirium trial routinely requires thousands of participants—the scarcity of outcome-level evidence stems not from laziness, but from statistical reality. So the more honest statement is: AI has indeed made the "depth number" more accurate and more anticipatory—but the critical step from "more accurate number" to "better patient outcomes" remains untraveled by anyone.
## Three Takeaways You Can Carry With You
If you or a loved one is about to undergo general anesthesia, you need not memorize any AUROC values—just remember three things. First, awareness risk is not evenly distributed: emergency surgery, cesarean delivery, cardiac/thoracic surgery, use of neuromuscular blocking agents, and obesity are all named risk factors—and two-thirds of cases occur during induction and emergence. Vigilance during these two phases matters more than monitoring the number throughout the entire procedure. Second, depth monitoring functions more like a dashboard than autonomous driving; the largest trials repeatedly fail to demonstrate that watching the number translates into better outcomes. Third, for older adults, aggregated evidence leans toward reduced anesthetic dosing under EEG guidance—but the two large North American trials failed to replicate this benefit. This remains an actively contested issue—worth discussing directly with your anesthesia team about how they plan to use monitoring.
One question worth debating remains: If next-generation AI reduces the prediction error of the old number to near zero—while the link between that old number and patient outcomes remains just as unclear—should we keep refining the same number, or should we switch to a different one?
Source: ENGAGES-Canada randomized trial (*JAMA* 2024). https://doi.org/10.1001/jama.2024.8144