Hypotension prediction: advancing the debate or retreading old ground?

Understanding arterial blood pressure and its management is a key endeavour in peri-operative medicine, as we grapple with the fundamental question of what target values arterial pressure should or should not reach to ensure well tolerated peri-operative care. The statistical relationships between intra-operative hypotension (IOH) and adverse outcomes have been well documented in large, predominantly observational studies by multiple independent research groups.1,2 Following reports of the prognostic relevance of IOH, the European Society of Cardiology 2022 guidelines, endorsed by the European Society of Anaesthesiology and Intensive Care Medicine, recommended avoiding reduction in intra-operative mean arterial pressure (MAP) for no more than 10 minutes of more than 20% from baseline values, or 60 to 70 mmHg (Class 1 recommendation).3 However, recent interventional studies and a meta-analysis challenge these initial observations, suggesting a weaker association between IOH and poor outcome.4–6 Mitigating the exposure to IOH has been the subject of intense research in peri-operative medicine in the last decade. In 2018, Hatib et al.7 first reported the hypotension prediction index (HPI), a machine learning proprietary algorithm based on 23 features derived from the arterial pressure waveform to predict the occurrence of IOH. In practical terms, HPI predicts IOH up to 15 min before its occurrence, and proponents argue that this is sufficient to allow for pre-emptive treatment. Clinicians are alerted to an impending IOH event when the index exceeds 85, a threshold value chosen to maximise the sensitivity and specificity of IOH occurrence. In this context, ‘prediction’ may be nothing more than diagnosis of an activation of compensatory mechanisms, as baroreflex activation in response to hypotension changes the mechanical properties of the radial artery and the resultant arterial waveform that is detected by the HPI. Supporting the real-world use of HPI, studies have demonstrated its effectiveness in reducing the cumulative duration of IOH compared to standard care.8 Moreover, centres that regularly use HPI report very low median time-weighted average MAP less than 65 mmHg, although whether this is a direct effect of HPI is unknown.9 Arguments against the use of HPI technology include the possibility of unnecessary interventions and the potential of inducing worse outcomes by giving treatments that may in themselves be deleterious.6,10 The debate about HPI covers several domains with different relevance for frontline clinicians. The first domain concerns methodology and statistics and is the most complex. Relevant discussions regarding the methodology used for the validation of HPI have been raised.11 Backward validation, wherein false positives and negatives are defined according to the occurrence of the event, then looking backward to determine if MAP criteria were fulfilled, provides more optimistic performance statistics. Forward validation, where false positives and negatives are defined according to the presence of MAP criteria, then looking forwards to see if an ‘alarm’ is set off, typically yields poorer predictive performance.12 Forward validation is more consistent with real-world practice and should be the method of choice. Regardless of methodological considerations, studies consistently show that HPI is able to detect IOH. The contentious issue is whether or not IOH may be reliably predicted by nonproprietary variables such as MAP alone. MAP-based methods appear to predict IOH,12–14 and recent studies show that a MAP threshold of less than 72 mmHg (MAP <72) predicts IOH as well as the HPI.12,14 HPI and MAP correlate highly, as MAP is one of the features used in the HPI algorithm, and times from alert to event are similar for HPI and MAP less than 72 mmHg.14 This has significant implications, given that a decision-threshold based on MAP less than 72 mmHg is freely available, whereas HPI is a proprietary algorithm packaged in a product that substantially increases costs. The second domain concerns a more practical approach, the management of false positives and false negatives of IOH. Readers should be aware that although the predictive performance of both HPI and MAP less than 72 mmHg are quite high (AUC approx.0.90), positive predictive values (PPVs) remain poor, around 30%, and are related to the prevalence of hypotensive events. In this regard, analysis using precision-recall curves that plot the PPV (precision) versus the corresponding sensitivity (recall) advocated by Mukkamala et al.13 may be a sensible approach as it indicates the average PPV across a range of sensitivity thresholds. Translated into practical terms, this would mean that there are many false positives, and clinicians should NOT immediately correct IOH ‘predicted’ either by high HPI or by MAP less than 72 mmHg values. Readers may wonder why the European Journal of Anaesthesiology has chosen to publish yet another article on the HPI, one that is a single-centre retrospective, exploratory re-analysis of a previous investigator-initiated prospective feasibility study.15 The findings, arguably, do not add substantially to current knowledge given recent publications demonstrating that an absolute MAP threshold of approximately 72 mmHg performs comparably to the HPI.12,14 Rellum et al.15 compared the HPI with various MAP thresholds for the prediction of IOH using three clinically relevant metrics: time from alert to event, PPV and proportion of missed hypotensive events. To their credit, the authors conducted extensive analyses comparing multiple HPI alerts against multiple MAP thresholds. However, readers may find the article difficult to dissect, as the different methods of comparing HPI with MAP, and the methods for classifying false negative may be a source of confusion. The final message proposed by the authors is that the HPI-85 pop-up alert, when the HPI > 85 is present for at least 40 s, is a more effective predictor of IOH than a MAP less than 72 mmHg threshold without a > 40 s alert, due to its better positive predictive value (55% for HPI-85 pop-up versus 30% for MAP <72 mmHg). The HPI-85 pop-up also demonstrated marginally increased time-to-event of 0.58 min, compared with MAP less than 72 mmHg. However, this is a difference that is unlikely to be of clinical significance and the proportions of missed hypotensive events were similar (in the 1 to 3% range) for HPI-85 popup and MAP less than 72 mmHg. Interestingly, when the HPI-85 was evaluated against MAP less than 72 mmHg without the requirement of the more than 40 s alarm (HPI-85 without the pop-up alert), the performances were comparable with equally poor PPV (approximately 30%). When a comparison between the two methods was conducted by requiring alert definitions for both HPI and MAP to persist for at least 40 s, their predictive performances were comparable. This makes sense, as the proportion of false positives would be expected to be reduced by filtering out short fluctuations in MAP. The authors should be congratulated for their extensive analyses; however, several issues remain unresolved. Firstly, the authors contend that the HPI prevents potential overtreatment, as MAP-alerts tend to trigger substantially more false predictions than HPI. This may be true but was not investigated, so the extent to which these methods may lead to overtreatment is unknown. Furthermore, the authors contend that the HPI-affiliated platform still offers additional variables for diagnostic guidance. Indeed, the HPI provides several useful secondary screens that may help users ascertain the mechanisms of the underlying physiological perturbations, but that was not evaluated in the present manuscript. Although the PPVs and NPVs are presented, the sensitivities and specificities are not directly presented making the estimation of likelihood ratios for positive and negative tests to assess the usefulness, or otherwise, of the tests, somewhat problematic. We agree with the author's contention that a MAP less than 72 mmHg threshold for predicting future IOH events should be subject to the same scrutiny as HPI. Beyond prediction of IOH, an added conundrum is the fact that BP control has not been shown to improve clinically important outcomes in randomised trials.4,5 Whether management of IOH (prevention versus correction), regardless of the method used, actually leads to better patient outcomes remains a top research priority. Beyond academic debates, the authors of this editorial would like to summarise a few take-home messages for frontline clinicians: (1) In light of recent publications, it seems that the statistical association between IOH and poor outcome is less robust than previously thought, emphasising a nuanced approach to IOH management. (2) Recent guidelines (ESC/ESAIC 2022) still recommend ‘avoidance’ of MAP decreases lasting longer than 10 minutes of more than 20% from baseline values or 60 to 70 mmHg. In practical terms, except for situations of severe IOH (MAP < 50 mmHg), clinicians have time to analyse the situation and provide adequate management. This message is even more important when the low PPVs discussed above are taken into consideration. (3) Avoidance of IOH (as opposed to simple indiscriminate correction through prescription of vasopressors) implies the application of a diagnostic tool for IOH. If clinicians choose to use the HPI technology (with the associated secondary screen to aid decision making), the monitor will, by default, activate the alarm when HPI is more than 85. If the clinicians choose to use more easily accessible (and cheaper) technologies such as MAP less than 70 to 75 mmHg, it is necessary to set appropriate alarm values. In conclusion, although the current study contributes valuable insights to the ongoing HPI debate, many questions remain unanswered, particularly regarding the clinical impact of IOH management on patient outcomes. Meanwhile, the EJA welcomes transparent discussions on the value of any peri-operative monitoring tool. We commend the authors on submitting their article to the EJA for scrutiny, and gratefully acknowledge the unseen work of the reviewers in providing honest, constructive and scientifically robust discussions surrounding this manuscript.