Implementation and learning curve in AI-assisted fluid management during abdominal oncologic surgery: a retrospective observational study

Abstract

Background

Intraoperative fluid management during major abdominal oncologic surgery is complex and highly operator-dependent. Assisted Fluid Management (AFM) is an artificial intelligence–based decision support system designed to guide fluid challenges based on real-time Stroke Volume (SV) analysis. However, limited data are available on how AFM is adopted in routine clinical practice and how clinician interaction with the system evolves over time.

Methods

We conducted a retrospective observational study based on a prospectively maintained institutional database at a high-volume tertiary referral center. Adult patients undergoing major abdominal oncologic surgery with intraoperative AFM monitoring were included. Two consecutive time periods following AFM implementation were compared. Analyses were performed at the fluid-challenge level and focused on patterns of fluid challenge initiation (clinician-initiated vs AFM-suggested), hemodynamic effectiveness (SV response), and bolus characteristics, as markers of system adoption and learning curve. Postoperative clinical outcomes were not assessed.

Results

Fifty-nine patients were included, accounting for 404 fluid challenges. Over time, clinician-initiated boluses significantly decreased and AFM-suggested fluid challenges increased (p < 0.001). This shift was associated with higher overall effectiveness of fluid challenges and greater SV responses, particularly for AFM-suggested boluses, which showed a significant improvement in effectiveness and ΔSV over time (p < 0.05).

Conclusions

Progressive integration of AFM into routine anesthetic practice was associated with measurable changes in clinician behavior and improved physiological effectiveness of intraoperative fluid challenges over time, consistent with a learning curve effect. These findings support the role of AI-based decision support systems in promoting more consistent and physiologically targeted fluid management and provide a foundation for future prospective studies evaluating their impact on clinical outcomes.

Introduction

Intraoperative hemodynamic instability remains a frequent challenge during major abdominal oncologic surgery and is associated with an increased risk of perioperative complications, particularly in high-risk patients. Fluid administration plays a central role in maintaining adequate cardiac output and tissue perfusion; however, both insufficient and excessive fluid loading may negatively affect organ function and surgical outcomes. Achieving an appropriate balance between these extremes requires continuous interpretation of dynamic hemodynamic variables and timely clinical decision-making [1–6].

Clinical decision support systems based on artificial intelligence are increasingly being welcomed by healthcare systems under significant pressure, with the aim of improving efficiency and reducing costs [7]. However, patients and clinicians often express greater apprehension toward the use of AI in healthcare [7, 8]. Human factors may play an important role in this process: initial skepticism or limited trust in AI-generated recommendations could reduce adherence to decision support outputs, particularly during the early phases of implementation [7, 8]. Over time, increasing familiarity with the system and experiential validation of algorithm performance may contribute to progressive changes in clinician behavior, potentially giving rise to a learning curve that may influence the functional effectiveness of AI-assisted interventions [8].

The Acumen Assisted Fluid Management (AFM) system (Edwards Lifesciences, Irvine, CA, USA) is a real-time clinical decision support tool designed to assist anesthesiologists in the administration of intraoperative fluid challenges [9–11]. By continuously analyzing Stroke Volume (SV) and its response to previous boluses, AFM provides recommendations on when a fluid challenge is likely to be effective, supporting a more targeted and physiologically driven approach to fluid administration [9, 11]. Unlike closed-loop systems, AFM operates as an open-loop decision support tool, leaving final therapeutic decisions to the clinician. [9–11]. Clinical trials and post hoc analyses have shown that AFM-guided fluid challenges are more likely to result in a meaningful SV increase compared with clinician-initiated boluses, both in major abdominal and high-risk surgical settings [12–14].

As with other decision support technologies, the clinical effectiveness of AFM does not rely solely on the underlying algorithm but also on how its recommendations are interpreted and integrated into routine practice. Familiarity with the system, confidence in its suggestions, and the ability to contextualize its outputs within the broader clinical picture are likely to evolve over time. This process implies the existence of a learning curve, during which clinicians progressively modify their behavior, potentially increasing adherence to AFM recommendations and refining the efficiency of fluid challenges [11].

The primary aim of this study was therefore to evaluate the clinical implementation of AFM over time in a high-volume tertiary referral center for major abdominal oncologic surgery. Specifically, we aimed to characterize temporal changes in clinician interaction with the AFM system, focusing on patterns of fluid challenge initiation (clinician-initiated versus AFM-suggested) and on the physiological effectiveness of administered boluses, as assessed by SV response. These implementation and performance metrics were used as indirect markers of a learning curve in the use of this AI-based decision support system. The study was not designed to assess postoperative clinical outcomes.

Methods

Study design and ethical approval

This study was conducted at the National Cancer Institute IRCCS “Fondazione G. Pascale”, Naples, Italy, a high-volume tertiary referral center for major abdominal oncologic surgery. It was designed as a retrospective observational analysis of a prospectively maintained institutional database.

The study was conducted in accordance with the Declaration of Helsinki and applicable local regulations. in accordance with the Declaration of Helsinki and applicable local regulations. Approval was obtained from the Institutional Ethics Committee on 27 May 2024 (Protocol No. 33/24).

Given the retrospective and non-interventional nature of the study, no study-specific procedures were performed. Written informed consent for data processing was obtained. For patients who were non-contactable or deceased, data were processed following a Data Protection Impact Assessment (DPIA) in accordance with the General Data Protection Regulation (GDPR) and Article 110-bis of the Italian Privacy Code. All data were anonymized prior to analysis.

Study population

All adult patients (≥ 18 years) who underwent major abdominal oncologic surgery between February 2024 and March 2025 and received intraoperative invasive arterial pressure monitoring connected to AFM were eligible for inclusion. For each included patient, baseline demographic and clinical characteristics were collected from the institutional database. These included age, sex, body mass index when available, American Society of Anesthesiologists (ASA) physical status, and type of surgical procedure.

Major abdominal oncologic surgery was defined as elective oncologic procedures involving the abdominal cavity, including colorectal resections, gastric surgery, hepatobiliary surgery, and cytoreductive surgery. Procedures were not selected based on duration thresholds; inclusion was determined solely by surgical type and availability of AFM-integrated hemodynamic monitoring.

Patients were excluded if they met any of the following criteria:

• Documented cardiac arrhythmias interfering with SV analysis;

• Acute myocardial ischemia, defined as clinically diagnosed acute coronary syndrome and/or ischemic electrocardiographic changes with associated biomarker elevation;

• Sepsis, defined according to Sepsis-3 criteria;

• Renal failure, defined as chronic kidney disease stage 4–5 (estimated glomerular filtration rate < 30 mL/min/1.73 m²) or need for renal replacement therapy;

• Emergency surgery;

• Pregnancy;

• Absence of an arterial catheter suitable for AFM monitoring.

To evaluate changes over time in the clinical use of AFM, patients were stratified into two temporal cohorts based on the phase of AFM implementation within the same anesthesia team:

• Group 1: February 2024–June 2024

• Group 2: July 2024–March 2025

These groups represent two consecutive implementation periods and do not constitute different treatment arms.

Anesthesia team and AFM implementation

Intraoperative anesthetic management was provided by a dedicated team of four attending anesthesiologists with expertise in major abdominal oncologic surgery. The same core group of attending anesthesiologists was involved throughout the study period, with substantial overlap between the two observation phases.

Prior to the start of the first observation period (February 2024), all team members received structured training on the use of the AFM system, including manufacturer-led in-service sessions, internal education on advanced hemodynamic monitoring, and supervised use during initial clinical cases. Following this implementation phase, AFM was progressively integrated into routine clinical practice without additional formal retraining.

No changes in institutional anesthetic protocols, hemodynamic targets, AFM software versions, or predefined system settings were introduced between the two study periods. Differences observed over time therefore reflect increasing familiarity with the system, growing confidence in AFM-generated recommendations, and progressive integration of decision support into clinical decision-making, consistent with a learning curve effect rather than protocol-driven changes.

Anesthetic and fluid management

Upon arrival in the operating room, a large-bore peripheral venous catheter was placed, and standard monitoring was applied, including a 5-lead electrocardiogram, pulse oximetry, and non-invasive blood pressure monitoring. When clinically indicated, a thoracic epidural catheter (T10–T12) was positioned for perioperative analgesia. General anesthesia was induced with propofol, fentanyl, and rocuronium bromide and maintained with volatile agents (sevofluorane or desfluorane), targeting a bispectral index (BIS) value between 40 and 50. Supplemental intravenous fentanyl was administered as required.

Mechanical ventilation was provided using a tidal volume of approximately 8 mL·kg⁻¹ of predicted body weight, with a positive end-expiratory pressure of 5 cmH₂O. The fraction of inspired oxygen was adjusted to maintain arterial oxygen saturation ≥ 96%, and respiratory rate was regulated to achieve an end-tidal CO₂ between 35 and 40 mmHg. After induction of anesthesia, an arterial catheter was placed in the radial artery in all patients for continuous invasive blood pressure monitoring.

A continuous basal infusion of balanced crystalloid solution was administered throughout surgery at a rate of 1–3 mL·kg⁻¹·h⁻¹ for laparoscopic or robotic procedures and 3–5 mL·kg⁻¹·h⁻¹ for open surgeries, in accordance with institutional practice. No changes to anesthetic management or maintenance fluid protocols were introduced between the two study periods.

Hemodynamic monitoring was performed using Acumen IQ™ with AFM sensor connected to the HemoSphere™ platform (Edwards Lifesciences, Irvine, CA, USA). The arterial pressure waveform was continuously acquired at a sampling rate of 100 Hz, and derived hemodynamic parameters were updated every 20 s. Displayed variables included systolic, diastolic, and mean arterial pressure, heart rate, SV, Stroke Volume Variation (SVV), pulse pressure variation (PPV), cardiac output, cardiac index, Hypotension Prediction Index (HPI), and AFM outputs when enabled.

Vasopressor and other cardiovascular medication use followed the institutional hemodynamic management protocol (Fig. 1) and was not modified for the purposes of this study. When indicated, vasopressors or inotropic agents were initiated and titrated at the discretion of the attending anesthesiologist based on the overall hemodynamic profile. This approach remained consistent across both study periods. Detailed temporal relationships between vasopressor administration and individual fluid challenges were not systematically captured and were therefore not included in the analysis.

Fig. 1.

Institutional hemodynamic management algorithm integrating AFM-guided FCs. When hemodynamic instability is detected, defined by Hypotension Prediction Index (HPI) ≥ 85 or mean arterial pressure (MAP) < 65 mmHg, patients are reassessed according to the institutional protocol. If an AFM-suggested fluid bolus is generated, fluid responsiveness is evaluated and managed accordingly. In the absence of an AFM suggestion, additional hemodynamic parameters—including peak rate of arterial pressure rise (dP/dt_max) and dynamic arterial elastance (Eddy)—are used to guide therapy. Depending on the underlying mechanism of instability, treatment options include fluid bolus administration, vasopressor therapy, inotropic support, or a combination thereof. After each intervention, hemodynamic reassessment is performed. This institutional protocol was applied consistently across both study periods

Fluid challenge strategy and AFM settings

Before surgery, clinicians selected a fluid strategy corresponding to a target increase in SV of 10%, 15%, or 20%, reflecting different degrees of stringency in defining fluid responsiveness. The threshold could be adjusted during the procedure according to clinical needs, such as surgical complexity, blood loss, major fluid shifts, or the need for a more restrictive or liberal fluid management strategy.

The selected SV threshold was used exclusively as a classification criterion to define the effectiveness of each fluid challenge, whether initiated by the clinician or suggested by the AFM system. These thresholds do not represent average or expected SV responses and do not define additional patient groups.

According to institutional practice and AFM system configuration, a reference fluid challenge volume of 250 mL was used to define and interpret fluid responsiveness. However, bolus volume and infusion rate were not rigidly protocolized and could vary within a clinically meaningful range (minimum 100 mL and maximum 500 mL), based on intraoperative conditions and clinician judgment. Fluid challenges were delivered manually, with an infusion rate sufficient to allow AFM analysis of the hemodynamic response, without a predefined fixed rate. These practices remained unchanged across the two study periods.

The AFM system can be set to target an SV increase of 10%, 15%, or 20% for a reference fluid challenge volume of 500 mL. When fluid challenges—initiated either directly by the clinician or suggested by the AFM system—of smaller volume are administered, the software automatically applies an internal scaling factor to correct for differences in delivered fluid volume. The relationship between SV change and fluid challenge volume is non-linear; for example, the scaling factor for a 250 mL fluid challenge is 1.467. When correcting for fluid volume using this constant, the AFM considers an SV increase ≥ 6.8% (i.e., 10%/1.467) as a positive response for its adaptive component (bolus log model). This normalization enables consistent assessment of fluid responsiveness across a range of clinically used bolus volumes, independently of the actual volume delivered [21]. No changes in AFM software settings or predefined thresholds were introduced between the two observation periods.

Data collection

For each patient, the following intraoperative variables were collected:

• Number of AFM-suggested and clinician-initiated fluid challenges;

• Bolus volume and resulting change in SV (ΔSV);

• Proportion of effective fluid challenges, defined as those achieving the pre-set SV response threshold.

Statistical analysis

Qualitative variables were presented using counts and percentage frequencies, and comparisons were conducted with the chi-square test. Quantitative variables were presented as median and interquartile range, or as mean and standard deviation when normally distributed. The normality of continuous variables was assessed using the Shapiro–Wilk test. The t-test or Wilcoxon rank-sum test was used for comparison of quantitative variables among the groups, as appropriate.

All analyses were conducted using STATA IC 15.1 (Stata Corp) for Windows, Microsoft Excel, R Foundation for Statistical Computing (Vienna, Austria, URL, https://www.R-project.org/), and the Acumen Analytics software (Edwards Lifesciences, Irvine, CA). A p-value of < 0.05 was considered significant.

Results

A total of 59 patients undergoing major abdominal procedures were included and divided into two temporal cohorts based on the implementation phase of the AFM system: 28 patients during the first period (group 1) and 31 patients during the subsequent phase (group 2).

Baseline demographic and procedural characteristics of the overall study population are reported in Table 1. The overall cohort had a mean age of 62.0 ± 12.8 years, and 49.2% were male. Most procedures were performed through an open surgical approach (67.8%), with a mean duration of surgery of 396.6 ± 206.5 min. All patients had an ASA physical status recorded, with the majority classified as ASA III. The study population included major abdominal oncologic procedures.

Table 1.

Baseline demographic and procedural characteristics of the study population

Variable	Overall cohort (n = 59)
Age, years (mean ± SD)	62.0 ± 12.8
Male sex, n (%)	29 (49.2)
Surgical approach–open, n (%)	40 (67.8)
Duration of surgery, min (mean ± SD)	396.6 ± 206.5
ASA physical status, n (%)
ASA II	6 (10.2)
ASA III	52 (88.1)
ASA IV	1 (1.7)
Type of surgery, n (%)
CRS–HIPEC	7 (11.9)
Hepatobiliary surgery	8 (13.6)
Gynecologic oncologic surgery	9 (15.3)
Colorectal surgery	5 (8.5)
Other complex abdominal oncologic surgery	30 (50.8)

Across the entire study population, fluid challenges were administered repeatedly within individual patients. The median number of fluid challenges per patient was 8 [IQR 4.5–12], with a wide dispersion reflecting differences in surgical complexity and intraoperative hemodynamic course. All subsequent analyses were therefore conducted at the fluid challenge level rather than at the patient level.

During the first observation period (28 cases, Table 2), a total of 206 fluid challenges (FCs) were administered. FCs were effective in increasing SV in 54% of cases, with a mean ΔSV of 15 ± 10% and an average bolus volume of 181 ± 86 mL. Of the total boluses, 47 were initiated directly by the clinician, whereas 68 were delivered following AFM recommendations; the remaining boluses corresponded to AFM test–boluses, generated when available information was insufficient to predict fluid responsiveness. The effectiveness of AFM-guided boluses in this phase was 59%, while clinician-initiated boluses were effective in 51% of cases.

Table 2.

Characteristics of fluid challenges (FCs) in the period February–June 2024 according to fluid strategy and fluid challenge initiation (clinician-initiated vs AFM-suggested)

	All strategies	10% strategy	15% strategy	20% strategy
All fluid boluses (n)	206	28	148	30
Volume (ml)	181 ± 86	192 ± 106	182 ± 82	167 ± 87
Effective FCs (n, %)	112 (54)	20 (71)	79 (53)	13 (43)
ΔSV (%)	15 ± 10	15 ± 10	15 ± 10	16 ± 9
Test dose (n)	91	6	69	16
Recommended (n)	68	18	46	4
Volume (ml)	178 ± 87	180 ± 102	181 ± 83	129 ± 49
Effective FCs (n, %)	40 (59)	14 (78)	24 (52)	2 (50)
ΔSV (%)	15 ± 10	17 ± 9	14 ± 10	14 ± 13
Clinician (n)	47	4	33	10
Volume (ml)	209 ± 90	198 ± 157	212 ± 86	202 ± 82
Effective FCs (n, %)	24 (51)	2 (50)	18 (55)	4 (40)
ΔSV (%)	15 ± 10	13 ± 14	15 ± 10	17 ± 8

ΔSV indicates the observed percentage change in SV following fluid bolus administration. The 10%, 15%, or 20% threshold reflects the setting selected by the attending anesthesiologist at the start of surgery and represents the chosen fluid management strategy. The selected threshold was used to evaluate the effectiveness of each fluid challenge, whether clinician-initiated or AFM-suggested

Data are presented as mean ± SD or number (percentage), as appropriate

During the second observation period (31 cases, Table 3), a total of 198 FCs were administered. The overall effectiveness of FCs increased to 64%, with a mean SV increase of 17 ± 9% and an average bolus volume of 215 ± 103 mL. The number of boluses initiated by clinicians decreased from 47 in the first period to 16 in the second (p < 0.001), while maintaining similar effectiveness and comparable mean SV increase (15 ± 10% vs 14 ± 11%) (Fig. 2). Conversely, the number of AFM-suggested fluid challenges increased from 68 in the first period to 129 in the second period (Fig. 3). Over the same time frame, AFM-suggested fluid challenges were more frequently associated with an effective SV response, increasing from 59% in the first period to 70% in the second, and with higher observed ΔSV values (15 ± 10% vs 18 ± 9%) (Fig. 4).

Table 3.

Characteristics of fluid challenges (FCs) in the period July 2024—February 2025 according to hemodynamic strategy and FCs initiation (clinician-initiated vs AFM-suggested)

	All strategies	10% strategy	15% strategy	20% strategy
All fluid boluses (n)	198	0	181	17
Volume (ml)	215 ± 103	NA	221 ± 106	160 ± 40
Effective FCs (n, %)	127 (64)	NA	116 (64)	11 (65)
ΔSV (%)	17 ± 9	NA	17 ± 9	19 ± 9
Test dose (n)	53	0	48	5
Recommended (n)	129	0	120	9
Volume (ml)	223 ± 101	NA	227 ± 103	168 ± 35
Effective FCs (n, %)	90 (70)	NA	82 (68)	8 (89)
ΔSV (%)	18 ± 9	NA	18 ± 9	22 ± 8
Clinician (n)	16	0	13	3
Volume (ml)	226 ± 111	NA	225 ± 123	233 ± 29
Effective FCs (n, %)	8 (50)	NA	7 (54)	1 (33)
ΔSV (%)	14 ± 11	NA	15 ± 11	10 ± 13

ΔSV indicates the observed percentage change in SV following fluid bolus administration. The 10%, 15%, or 20% threshold reflects the setting selected by the attending anesthesiologist at the start of surgery and represents the chosen fluid management strategy. The selected threshold was used to evaluate the effectiveness of each fluid challenge (FC), whether clinician-initiated or AFM-suggested

Data are presented as mean ± SD or number (percentage), as appropriate

Fig. 2.

Comparison between two consecutive periods in the administration of FCs initiated by the clinician. Left: Total number of fluid boluses initiated directly by the clinician during the first study period (February–June 2024) and the second study period (July 2024–February 2025). Right: Proportion of positive hemodynamic responses to clinician-initiated boluses during the two study periods

Fig. 3.

AFM-suggested FCs and hemodynamic effectiveness across the two study periods. Left: Total number of fluid challenges suggested by the AFM system during the first study period (February–June 2024) and the second study period (July 2024–February 2025). Right: Proportion of AFM-suggested fluid challenges associated with an effective hemodynamic response during the two study periods, illustrating temporal changes in AFM-guided fluid administration

Fig. 4.

Comparison between clinician-initiated fluid boluses (Manual) and AFM-suggested fluid boluses (AFM) during the second study period (July 2024–February 2025). A Distribution of fluid boluses achieving a positive hemodynamic response in the two groups. B Distribution of SV increase following fluid bolus administration in the two groups

A comprehensive comparison of FC characteristics between the two study periods according to fluid challenge initiation is reported in Table 4.

Table 4.

Comparison of FC characteristics between the two study periods according to the source of FC initiation (clinician-initiated vs AFM-suggested)

Parameter	Group 1 (n 28)	Group 2 (n 31)	p value
Total FCs (n)	206	198	–
Effective FCs (n, %)	112 (54)	127 (64)	ns (Chi2)
ΔSV (%)	15 ± 10	17 ± 9	ns (t-test)
Bolus volume (ml)	181 ± 86	215 ± 103	ns (t-test)
Clinician-initiated FCs	47	16	< 0.001 (Chi2)
Effective FCs (n, %)	24 (51)	8 (50)	ns (Chi2)
ΔSV (%)	15 ± 10	14 ± 11	ns (t-test)
AFM-suggested FCs	68	129	< 0.001 (Chi2)
Effective FCs (n, %)	40 (59)	90 (70)	0.01 (Chi2)
ΔSV (%)	15 ± 10	18 ± 9	0.03 (t-test)

ΔSV indicates the observed percentage change in SV following fluid bolus administration. The 10%, 15%, or 20% threshold reflects the setting selected by the attending anesthesiologist at the start of surgery and represents the chosen fluid management strategy. The selected threshold was used to evaluate the effectiveness of each FC, whether clinician-initiated or AFM-suggested

Data are presented as mean ± standard deviation or number (percentage), as appropriate

Discussion

The clinical adoption of artificial intelligence tools requires a gradual learning process among healthcare professionals, during which familiarity, practical competence, and trust in the algorithm’s reliability are progressively developed.

In this context, our study provides a real-world evaluation of the implementation of AFM in a tertiary referral cancer center, focusing on temporal changes in clinician behavior and functional hemodynamic responses rather than on clinical outcomes.

Across the two observation periods, we observed a shift in intraoperative fluid management practice [15]. Specifically, the second period was characterized by a reduction in clinician-initiated fluid challenges and a parallel increase in AFM-suggested boluses. This transition was associated with higher overall effectiveness of fluid challenges and greater SV responses, suggesting a more selective and targeted use of fluids. Together, these findings are consistent with a learning curve effect, whereby increasing familiarity with AFM and growing confidence in its recommendations translate into changes in clinical decision-making and improved physiological efficiency of fluid administration.

Michard et al. emphasize that specific training on both the capabilities and limitations of AI—beginning as early as undergraduate education—is essential not only to dispel the perception of AI as an opaque “black box,” but also to empower clinicians to interpret the data provided, integrate it into the clinical context, and make informed decisions [16]. This educational pathway underpins the learning curve that enables a transition from initial skepticism or superficial use to meaningful clinical integration of AI systems.

Recent evidence supports the existence of a learning curve in the adoption of AI-based technologies. For example, a multicenter study conducted within the UK National Health Service (NHS) demonstrated that increased institutional experience with computed tomography (CT)-derived fractional flow reserve (FFR-CT) technology led to significant improvements in diagnostic accuracy, care pathway efficiency, and reductions in unnecessary investigations—underscoring the need for a progressive and structured process of clinical integration [17]. Similarly, our findings suggest that AFM effectiveness is not solely related to the algorithm itself, but also on how its outputs are interpreted and applied by clinicians over time.

Implementing a fluid therapy pathway supported by the AFM system within a healthcare institution is not a trivial undertaking and involves upfront costs. For successful adoption, a clear commitment to change is required—both at the clinical and administrative levels. One of the most critical success factors is the presence of a dedicated anesthesia team tasked with overseeing implementation. At our institution, AFM implementation was supported by a group of anesthesiologists specialized in major abdominal surgery, who underwent advanced hemodynamic training and developed internal protocols, potentially facilitating progressive integration of the system into routine practice.

When interpreted in the context of existing literature, our results are broadly consistent with prior studies evaluating AFM and other decision support systems, including the MicroSupport Trial and the work by Coeckelenbergh et al., which demonstrated the potential of algorithm-guided fluid challenges to improve the physiological effectiveness of fluid administration [12–14].

However, unlike randomized trials designed to test predefined therapeutic strategies, the present study specifically addresses the process of implementation and clinician interaction with AFM in everyday practice.

Several implications arise from these findings. First, they suggest that increasing exposure to AFM may promote more consistent and selective fluid administration, potentially reducing unnecessary boluses while preserving hemodynamic effectiveness. Second, they highlight the importance of organizational factors—such as dedicated teams, training, and institutional commitment—in determining the success of AI-based tools. Finally, they provide a rationale for future prospective studies designed to assess whether these functional improvements in intraoperative hemodynamic management translate into better postoperative outcomes.

The experience gained from AI-driven fluid management paves the way for future research and clinical implementation. A crucial next step will be to validate the impact of AFM on clinically meaningful outcomes, such as the incidence of acute kidney injury, major postoperative complications, and hospital length of stay. Large prospective and multicenter studies will be essential to determine the generalizability of these findings and to define the role of AI-assisted decision support in complex surgical populations.

In this regard, our findings are aligned with the rationale of the ongoing PEFLA Trial, a large stepped-wedge, multicenter international study specifically designed to evaluate whether AFM-guided intraoperative fluid management can reduce major postoperative complications and mortality in high-risk abdominal surgery. This trial is expected to provide definitive evidence on the clinical benefits of decision support-guided fluid therapy [18].

Another strategic priority will be the development and implementation of multicentre studies aimed at testing the reproducibility and generalizability of AI-guided fluid strategies in high-risk surgery. Such studies will be essential to consolidate the evidence base for AFM and define its role in complex surgical settings characterized by marked hemodynamic variability.

Limitations

This study has several limitations. First, its retrospective observational design inherently limits causal inference and may be subject to selection bias and unmeasured confounding. Second, analyses were conducted at the level of individual fluid challenges rather than at the patient level; because multiple fluid challenges were administered within the same patient, within-patient clustering could have influenced the results and may limit the independence of observations. Third, the single-center setting and the specific organizational context may restrict the generalizability of the findings.

In addition, postoperative clinical outcomes and the duration of hypotensive episodes were not evaluated, as the study was not designed to assess these endpoints. Furthermore, detailed quantitative data on vasopressor use and precise temporal relationships between vasopressor administration and individual fluid challenges were not systematically captured and may represent a source of residual confounding.

Conclusions

In conclusion, our results suggest that the progressive integration of an AI-based decision support system such as AFM may be associated with changes in intraoperative fluid management practices over time, characterized by increased reliance on algorithm-suggested fluid challenges and improved physiological responses. These findings primarily reflect implementation patterns and functional hemodynamic metrics rather than clinical outcomes. By supporting, rather than replacing, clinician decision-making, AFM may facilitate more physiologically targeted fluid administration in high-risk abdominal oncologic surgery. Prospective, multicenter studies will be required to determine whether these implementation-related improvements translate into meaningful patient-level benefits across the perioperative care pathway.

Authors’ contributions

L.F. and G.P. wrote the main manuscript text; M.M., C.I., R.A. collected data; G.P., L.F., A.B. and P.S. performed literature research; F.C., V.P. and F.V. performed data analysis; A.C. supervised the project. All authors reviewed the manuscript.

Funding

Not applicable.

Data availability

No datasets were generated or analyzed during the current study.

Declarations

Ethics approval and consent to participate

The study was approved by the Institutional Ethics committee on 27th May 2024 (Prot. 33/24).

Consent for publication

All patients provided written informed consent. For those who were non-contactable or deceased, a Data Protection Impact Assessment (DPIA) was carried out in accordance with the GDPR and Article 110-bis of the Italian Privacy Code.

Competing interests

Dr L. Frassanito and Dr G. Pasta have received funding from Edwards Life sciences (Irvine, California, USA) for educational and teaching activities. The remaining authors declare no competing interests. These activities were unrelated to the present study, and the manufacturer had no role in the study design, data collection, data analysis, manuscript preparation, or decision to submit the manuscript.