Artificial Intelligence for Hemodynamic Monitoring: The Art, Science, and the Machine!

Prediction of hypotension

It is due to the rampant technological advancements that the patterns and/or associations in the peculiarly large datasets can be effectively tracked in the present era. Motivated by the same, a logistic regression model was employed by Hatib et al.^[14] to devise the hypotension prediction index (HPI). It involves a machine learning (ML) algorithm premised on the high-fidelity analysis of thousands of arterial pressure waveforms to be subsequently expressed as a unitless numerical value between 0 and 100, denoting the HPI value.^[14] Herein, the Acumen^™ HPI software (Edwards Lifesciences; Irvine, CA, USA) alerts the physician of the likelihood of the patient experiencing a hypotensive event, i.e., mean arterial pressure (MAP) below 65 mmHg for at least a minute.^[4,15,16] With an elevated HPI denoting a higher likelihood of hypotension, values in excess of 85 on the HemoSphere advanced monitoring platform trigger the acoustic-visual alarm, a pop-up alert window, alongside a secondary screen.^[15] The secondary screen displays other quantitative hemodynamic parameters such as the cardiac output (CO), stroke volume (SV), SV variation (SVV), dP/dt (the change of pressure over time to indicate the cardiac contractility), systemic vascular resistance (SVR), and the dynamic arterial elastance, to extend guidance on the use of inotropes, fluids, and vasopressors, etc.^[4,15,16]

Of note, the ML algorithm proposed by Hatib et al.^[14] went on to predict arterial hypotension 15 min before an event with an area under the curve (AUC) of 0.95 with sensitivity and specificity of 87.5% and 87.3% in the internal validation cohort. The respective AUC for 10 min and 5 min time-to-event emerged to be 0.95 and 0.97.^[14] In the external validation cohort, the corresponding predictive AUC for 15 min, 10 min, and 5 min before a hypotensive event was discovered to be 0.91, 0.92, and 0.95, respectively.^[14] Furthermore, encouraging literature has also been accruing on its use in major non-cardiac surgery and the cardiac surgical patient population in the form of “European multicenter prospective observational registry” and the “Hypotension Prediction 2” randomized trial.^[17,18] At the same time, the diagnostic utility of the non-invasive ClearSight system when combined with the HemoSphere monitoring platform is being increasingly evaluated across diverse operative settings.^[19-21]

Hemodynamic profiling

Aligned with the “omics” outlook of precision perioperative medicine,^[6,22] Kouz et al.^[23] outlined some six distinct endotypes of hypotension in the intraoperative period, having employed hierarchical clustering. The discovered endotypes linked to different etiologies ranging from depressed myocardial function to bradycardia to vasodilation and from hypovolemia to the mixed type, and hence associating with specific lines of therapeutic amelioration.^[6,23] Quite similarly, however, more recently, Jian et al.^[24] developed an unsupervised deep learning algorithm to characterize the hypotensive endotypes based on the heart rate, indexed SV and SVR, and the prevailing SVV. The allied data emanating from 871 surgical patients (6,962 hypotension events) was validated in independent data sets of 1,000 surgical (7904 hypotension events) and another 1,000 critically ill patients (53,821 hypotension events) with hypotension labeled as MAP <65 mmHg lasting for at least a minute.^[24]These research findings, put together, adequately highlight the importance of hemodynamic profiling in guiding the context-appropriate causal treatment.^[23,24]

The concept of hemodynamic profiling extends further into the intensive care unit (ICU). Geri et al.^[25] again utilizing hierarchical clustering, nonetheless with the combined clinical and echocardiography data, explored cardiovascular phenotypes in a cohort of 360 septic-shock patients. They could identify the following profiles: Well-resuscitated, persistently hypovolemic, hyperkinetic, right ventricular (RV) failure, and left ventricular (LV) systolic dysfunction.^[25]Buttressing the role of clustering approach,^[6,26] Guinot et al.,^[27] aided by the clinical, biological (N-terminal pro-B-type natriuretic peptide), and the Doppler echocardiographic data, decipher the congestive phenotypes in ICU patients: Hemodynamic congestion with normal cardiac function, those with moderate ventricular function alterations, and systemic congestion with severe biventricular function alterations.^[27] The phenotypes were found to be eventually significantly different with regard to the renal outcomes and mortality rate, alongside the ICU and hospital length of stay, reiterating the part hemodynamic profiling has to play in the critically ill.^[26,27]

Early warning institution

Real-time hemodynamic monitoring with AI and the instantaneous data processing with ML algorithms enhances the dynamic understanding of a patient’s existing cardiovascular status, providing for the function of an early warning system (EWS).^[28,29] In a preliminary single-center study of elective non-cardiac surgical patients by Wijnberge et al.,^[28] the application of ML-derived EWS resulted in reduced intraoperative hypotension. To add, Hyland et al.^[29] engineered CircEWS and CircEWS-lite, to alert the clinicians to an anticipated circulatory failure within 8 h. Having developed and validated the former systems in the backdrop of a high-resolution ICU database, a continuous risk score was generated and updated every 5 min to forecast the circulatory failure likelihood.^[29] The model could achieve 90% prediction for circulatory failure in the test set, with 82% of the events being identified more than 2 h before.^[29] This resulted in an area under the receiver operating characteristic curve and an area under the precision recall curve amounting to 0.94 and 0.63, respectively, with the system raising 0.05 alarms/individual patient and hour, on average.^[29]

Closed-loop and decision-support system

Given the high-stake perioperative environment, the integration of AI can render the decision-making framework dynamic and adaptive, not only through multimodal data incorporation but through closed-loop optimization, combining therapy with monitoring.^[30] This would undoubtedly represent the most efficient and intelligent use of AI in hemodynamic monitoring, however, not without identifying the parameters of interest, investigating available modalities, developing meaningful algorithms, validating the closed-loop systems, and demonstrating the clinical utility in well-designed trials.^[30,31] Moreover, a systematic review by Pereira et al.^[32] although suggesting the utility of clinical decision-support system (CDSS) among cardiac surgical patients in the ICU, it marks the need for improvement to enhance the overall applicability. Ahead of the potential advantages of automated decision-making, the challenges posed by data quality, interpretability, and interoperability need to be concomitantly borne in mind, especially when an effective delivery of therapies in the perioperative period happens to be the primary goal of CDSS and closed-loop systems.^[4,31]

Assisted ultrasonography

The scope of AI-assisted ultrasonography is being increasingly recognized, especially for enhancing usability among the users with varying degrees of clinical experience. As for hemodynamic monitoring, a study by Shaikh et al.^[33] enrolled 28 trainees with limited experience in point-of-care ultrasound (POCUS), to reveal the feasibility of AI use for image acquisition in critical care. With the automated velocity time integral (VTI) measurements in the trainee subset associated with higher reproducibility, the assisted computation hinted toward an improved accuracy for the quantification of CO.^[33] While ML algorithms have mostly been formulated for the real-time LV ejection fraction estimation, they are coming up for the automated VTI, inferior vena cava, and LV diastolic function assessment which can be missed on a quick POCUS scan with qualitative focus.^[26] Interestingly, a systematic review and meta-analysis published in 2026 concluded that AI models, despite an encouraging accuracy for the assessment of RV function, display considerable heterogeneity, necessitating meticulous interpretation and future research.^[34] With a total of 5 databases being searched here employing the MeSH and Emtree terms: “Artificial Intelligence,” “Right Ventricular Function,” and “Right Ventricular Dysfunction,” the discovered heterogeneity (I² = 71.63%) could be partly explained by the algorithm type and the study country.^[34]

Futuristic technology

Meanwhile, the innovative real-time adaptive models, in general, signify a trend toward efficient personalized care; the use of edge computing for wearable device technology surfaces as a noticeable advancement.^[4,6,35] Furthermore, the battery of hemodynamic information that can be noninvasively made available with the ML algorithms trained with photoplethysmographic waveforms is yet to be determined completely.^[26] Finally, the role of implantable systems, Cordella^™ and CardioMEMS^™, showcasing potential for post-discharge pulmonary artery pressure monitoring in chronic heart failure, also remains to be explored in the surgical settings.^[5,36,37]