A blood-pressure algorithm predicts surgical hypotension. A simple alarm matched it.

The Hypotension Prediction Index (HPI) is a machine-learning algorithm that reads features of the arterial blood-pressure waveform and outputs a 0–100 score meant to warn an anaesthetist that a patient's blood pressure is about to fall, minutes before it does [s2]. Three randomised trials published in 2026 reach the same uncomfortable conclusion: the algorithm does help keep blood pressure up during surgery, but a plain alarm set on the mean arterial pressure does just as well — and where HPI wins, it wins because clinicians treated more aggressively, not because the software saw further ahead [s1][s2][s3].

What the device is, and what it claims

HPI was authorised by the US Food and Drug Administration in March 2018 through the De Novo pathway, as the Acumen Hypotension Prediction Index Feature Software made by Edwards Lifesciences, and later expanded in two clearances [s5]. Its selling point is anticipation: rather than reacting once a patient is already hypotensive, the anaesthetist acts on a rising score. That framing matters clinically, because intraoperative hypotension has been associated, in observational data, with acute kidney injury, myocardial injury and death [s1].

The catch, visible in the trials' own background sections, is that the HPI score is strongly correlated with the mean arterial pressure (MAP) it is derived from, and its alerts tend to fire when MAP is already around 70–75 mmHg [s1][s3]. If a device that "predicts" a fall mostly triggers when pressure is already drifting down, the honest question is whether it beats simply watching the pressure. In 2026 three groups tested exactly that.

The head-to-head trials

The first trial randomised 143 adults having moderate- or high-risk elective non-cardiac surgery to either a MAP alarm set at below 72 mmHg or the default HPI alarm above 85, and asked whether the simple alarm was non-inferior [s1]. It was. The median area under the hypotension threshold (MAP below 65 mmHg) was 3.75 mmHg·min (interquartile range 0.00–22.62) with the MAP alarm and 4.00 mmHg·min (0.00–19.00) with HPI; after log-transformation the mean difference was 0.03 (95% CI −0.24 to 0.29), with the lower bound comfortably above the pre-specified non-inferiority margin of −0.4 [s1]. Nothing else — alarm duration, drugs, fluids, kidney or heart injury, 30-day mortality — differed [s1].

The second trial went further and tested superiority. It randomised 100 adults undergoing major non-cardiac surgery to HPI-guided treatment (triggered at a score of 85 or above) or to a proactive higher-MAP target (treat when MAP fell to 73 mmHg or below), with both arms following the same treatment protocol [s2]. HPI was not superior. Time-weighted average hypotension was 0.07 mmHg (0–0.20) with HPI against 0.16 mmHg (0.02–0.50) with the MAP target (P = 0.119), and the area under the threshold was 22 mmHg·min (0–96) against 59.7 mmHg·min (5.7–119.3) (P = 0.172) [s2]. Norepinephrine dose, length of stay and 30-day mortality did not differ [s2]. The authors are careful about what this shows: the trial was not designed to prove equivalence, so the result is "no demonstrated superiority," not proof the two are identical [s2].

Why earlier trials looked better

Plenty of earlier trials did report that HPI reduced hypotension, and a 2026 meta-analysis of 18 randomised trials in 2,279 non-cardiac patients (plus one trial of 130 cardiac patients) found HPI-guided care lowered the time-weighted average of hypotension by a weighted mean difference of −0.19 mmHg (95% CI −0.25 to −0.12) and cut fluid and phenylephrine use [s4]. But the same meta-analysis found no effect on the outcome that motivates the whole exercise: postoperative acute kidney injury, with an odds ratio of 0.83 (95% CI 0.64 to 1.07, P = 0.16) [s4].

A separate 2026 analysis explains the gap between the impressive process numbers and the flat clinical ones. Across 20 randomised HPI trials in 2,342 patients, it asked which trials also treated patients more intensively [s3]. Of the 13 trials that reduced hypotension, 9 — 69% — had significantly greater use of fluids, vasopressors or inotropes in the HPI arm; none of the 5 trials that failed to reduce hypotension showed increased treatment intensity [s3]. The association was statistically significant (P = 0.029), and where trials reported it, the size of the hypotension reduction tracked treatment intensity closely (P < 0.001) [s3]. In plain terms: the trials where HPI "worked" are largely the trials where clinicians, prompted by the score, pushed harder — something a MAP alarm at the same threshold would do just as well.

The trial that appears to break the pattern

Not every 2026 trial was null. A study of 171 women having caesarean sections under spinal anaesthesia found HPI-guided care produced a much lower time-weighted average of hypotension — 0.08 mmHg — than continuous non-invasive pressure monitoring (0.30) or standard intermittent cuff readings (0.89) (P < 0.001), with less nausea and vomiting [s6]. The result is real, but it illustrates the same point rather than contradicting it: HPI was compared against reactive monitoring, not against a matched proactive MAP target [s6]. The comparison that flatters the algorithm is the one that does not hold the treatment threshold constant.

What it means

None of this says the HPI score is fake or that anticipating hypotension is a bad idea. It says the specific commercial claim — that a proprietary machine-learning index delivers something a blood-pressure number cannot — is not supported when the comparison is fair [s1][s2][s3]. A MAP alarm at roughly 72 mmHg is cheap, already built into every monitor, and in these trials matched an added-cost algorithm [s1]. For a device that bears on how aggressively a patient is treated during surgery, that is the comparison that should drive purchasing and protocol decisions.

What to watch

Whether any HPI trial is powered for a hard outcome — kidney injury, myocardial injury, mortality — rather than for millimetres of mercury below a line, and whether it randomises against a proactive MAP target rather than reactive monitoring. Until then, the burden of proof sits with the algorithm, not the alarm.

Sources

• [s1] Hypotension Prediction Index versus Mean Arterial Pressure Alarm for Preventing Intraoperative Hypotension in Elective Non-Cardiac Surgery: A Randomized Controlled Trial — Anesthesiology, 2026-08-19
• [s2] Hypotension Prediction Index versus High Mean Arterial Pressure Target for Preventing Intraoperative Hypotension: A Randomized Controlled Trial — Anesthesiology, 2026-08-17
• [s3] Associations between reduction of hypotension and treatment intensity in hypotension prediction index studies — Anesthesiology, 2026-09-09
• [s4] Hypotension prediction index in surgical patients: A systematic review and meta-analysis of randomized controlled trials — Medicine, 2026-06-16
• [s5] De Novo classification DEN160044: Acumen Hypotension Prediction Index (HPI) Feature Software — U.S. Food and Drug Administration, 2018-03-16
• [s6] Proactive haemodynamic management using the hypotension prediction index during caesarean section: a randomised controlled study — Anaesthesia, 2026-02-15