A Novel Approach to Early Personalized Hemodynamic Resuscitation: Non‐Invasive Peripheral Photoplethysmography for Identifying Predominant Vasodilatory Shock in Sepsis

Authors: Sanne Ter Horst, Anna D. Schoonhoven, Raymond J. van Wijk, Rick Weitering, Sanne W. van Loon, Jan C. ter Maaten, Hjalmar R. Bouma

Abstract

Introduction

Sepsis remains a leading cause of mortality, with mortality from septic shock exceeding 40%. Standardized resuscitation (30 mL/kg) may cause adverse outcomes, including fluid overload or prolonged hypotension, emphasizing the need for individualized strategies. Sepsis‐induced shock arises from varying degrees of vasodilation and hypovolemia, yet patients often present with similar clinical signs in the emergency department (ED). Photoplethysmography (PPG), a non‐invasive technique reflecting peripheral perfusion, may help identify patients with a predominant vasodilatory profile who could benefit from early vasopressor therapy.

Methods

This post hoc analysis used data from the Acutelines biobank at the University Medical Centre Groningen. Adults admitted for non‐trauma specialties with suspected infection and hemodynamic instability (MAP < 70 mmHg, SBP < 90 mmHg, shock index > 0.9, or lactate > 4.0 mmol/L) were included. PPG data were pre‐processed and features extracted. Principal component analysis (PCA) and K‐means clustering enabled dimensionality reduction and hemodynamic profiling. Logistic regression assessed the discriminative performance of PPG‐based models for vasopressor therapy initiation within 24 h.

Results

Among 325 patients, 16.3% received vasopressors. PCA identified three principal components explaining 80.3% of PC1 (arterial compliance), PC2 (cardiac output and systemic vascular resistance), and PC3 (peripheral vasomotor tone). The PPG‐based model showed moderate discriminative power (AUROC: 0.75), improving when combined with MAP and lactate (AUROC: 0.83).

Conclusion

PPG enables identification of patients likely to benefit from vasopressor therapy during the first 20 min after ED arrival. By providing additional insight into peripheral perfusion, this proof‐of‐principle study supports further exploration of PPG as a clinical support tool for personalized hemodynamic resuscitation in sepsis.

Editorial Comment

This secondary analysis demonstrates early peripheral circulatory patterns in sepsis using photoplethysmography at the start of resuscitation. Distinct PPG‐derived profiles were associated with vasopressor initiation within 24 h, supporting PPG as a tool for personalized resuscitation.

Introduction

Sepsis is the leading global cause of mortality, with a mortality rate of 10% that increases to more than 40% in the case of septic shock [1, 2]. In 2017, the World Health Organization [WHO] estimated that sepsis affected 49 million people globally, resulting in 11 million deaths worldwide [3]. Sepsis is defined as a life‐threatening organ dysfunction caused by a dysregulated host response a reaction to infection, culminating in organ failure and shock [2]. Early treatment of hemodynamic instability in a patient with sepsis is of utmost importance to prevent organ failure and death.

Current sepsis treatment includes timely antibiotics and hemodynamic resuscitation with intravenous fluids and, when needed, vasopressor therapy [1, 4]. The Surviving Sepsis Campaign Guidelines still recommend administering at least 30 mL/kg of crystalloid fluids for patients with sepsis‐induced hypoperfusion or septic shock to restore organ perfusion [4, 5]. Vasopressors are indicated in cases of persistent hypotension despite adequate fluid resuscitation, or when fluids are either unlikely to restore hemodynamic stability or contraindicated, targeting a mean arterial pressure (MAP) of ≥ 65 mmHg [6]. However, a one‐size‐fits‐all approach is problematic due to significant inter‐individual differences in fluid responsiveness and tolerance [7, 8]. Only about half of patients are fluid responsive; others risk fluid overload [9, 10, 11], which can lead to venous congestion, worsened organ dysfunction, prolonged intensive care unit (ICU) stays, and increased mortality [12, 13, 14, 15]. Notably, signs of venous congestion have been observed in both fluid responsive and unresponsive critically ill patients, underscoring the complexity of fluid management [10]. Emerging evidence suggests that very early vasopressor support seems safe, may reduce the total volume of fluids required, and thereby could improve clinical outcomes [16, 17]. These findings emphasize the need for a personalized approach to hemodynamic resuscitation, carefully balancing the choice between higher‐volume fluid resuscitation or early initiation of vasopressor therapy to optimize patient outcomes [9, 18].

A personalized strategy based on patient‐specific cardiovascular changes is essential to optimize resuscitation, prevent organ failure, and potentially reduce sepsis‐related morbidity and mortality. Septic shock typically involves both vasodilatory and hypovolemic components, with some patients also experiencing cardiogenic shock due to septic cardiomyopathy [18, 19, 20]. Vasodilatory shock, marked by excessive vasodilation and low systemic vascular resistance, requires vasopressors to restore vascular tone [21, 22]. Hypovolemic shock can be absolute, caused by fluid loss from vomiting, diarrhea, or bleeding, or relative, resulting from venous vasodilation and increased venous capacitance that reduce venous return [22, 23]. While absolute hypovolemia primarily requires fluid administration, relative hypovolemia may benefit from both fluids and vasopressors to improve venous return and maintain perfusion [23, 24]. Complete differentiation between these mechanisms is often not possible, but identifying the predominant physiological driver can help guide early and effective management. Therefore, incorporating cardiovascular parameters that reflect these pathophysiological changes could improve individualized treatment decisions and help identify sepsis patients who would benefit from early vasopressor therapy.

Photoplethysmography (PPG) is a non‐invasive optical technique commonly integrated in pulse oximeters [25]. While pulse oximetry primarily measures oxygen saturation, PPG waveform analysis provides additional insight by reflecting peripheral blood volume changes [25, 26], which are influenced by cardiac output, arterial compliance, systemic vascular resistance, and vasomotor tone [27, 28]. We hypothesize that PPG‐derived features reflect the hemodynamic effects of sepsis, particularly peripheral vasodilation, enabling identification of patients who might benefit from early vasopressor therapy in the emergency department (ED). Therefore, in this proof‐of‐principle study, we first aimed to identify early septic shock subphenotypes based on PPG‐derived features. Second, we explored whether PPG measured in the first 20 min of ED triage could identify individual patients with a predominant vasodilatory shock component, as defined by receiving vasopressor therapy within 24 h, who may potentially benefit from early vasopressor intervention.

Methods

Study Design

We conducted a post hoc analysis of prospectively obtained data by Acutelines: a data‐, image‐, and bio‐bank at the ED of the University Medical Center Groningen (UMCG) [19]. A deferred consent procedure (by proxy) was in place to enable the collection of data and biomaterials before obtaining written consent. When reaching the patient or proxy was not feasible, we followed an opt‐out procedure. Acutelines is approved by the medical ethics board of the UMCG and registered under trial registration number NCT04615065 at ClinicalTrials.gov. The study protocol was approved by the institutional review board under registration number 18089. Acutelines' complete protocol and overview of the actual, entire data dictionary is available via www.acutelines.nl [29].

Study Population

The inclusion criteria for this study patients with an age above 18 years (i), admitted for either internal medicine, rheumatology, gastro‐enterology, pulmonology, or emergency medicine for a non‐traumatic reason (ii), who had a suspected infection based on physicians' discretion at ED arrival (iii), in need of hemodynamic resuscitation based on MAP < 70 mmHg, systolic blood pressure (SBP) < 90 mmHg, shock index (heart rate (HR)/SBP) > 0.9, or lactate > 4.0 mmol/L, and (iv) had available PPG waveform data captured by the Philips bed‐side monitor (v). The exclusion criteria patients with a non‐ICU admission policy outlined in an advanced care directive (i), and absent or low‐quality PPG data according to the determined signal quality index (SQI) described in the section below (ii). For the main study cohort, we included all eligible patients between September 2020 and December 2023. For the internal validation cohort, we included all eligible patients between January 2024 and June 2024.

Data Capture and Definitions

To allow collection of data and biomaterials upon first contact, primary screening of patients for eligibility upon arrival in the ED was performed 24/7 by the ED‐nurse with a trained research team. Continuous bedside physiological waveforms (e.g., electrocardiography (ECG), PPG, and impedance pneumography) and vital parameters were automatically captured by a Philips IntelliVue MP70 or MX550 with a multi‐measurement module bedside monitor in the ED and securely stored on a network drive for later analysis (Sept 2020–Dec 2021, Apr 2022–July 2022, and Jan 2023–June 2024). Information from other data sources such as the hospital's electronic health records was securely imported. Study data was collected using REDCap electronic data capture tools hosted at the UMCG [30]. For the main study cohort, we included all PPG and clinical data, including primary and secondary endpoints. For the internal validation cohort, we collected PPG data, age, gender, cardiovascular comorbidities, intravenous (I.V.) fluids, vital signs, and the use of vasopressor therapy < 24 h. For this study, we used the first 20 min of PPG waveform recorded after arrival at the ED, which matches the moment of triage when the patient's clinical condition is typically assessed based on vital parameters. Infection foci were confirmed post hoc based on the assessment of an expert adjudication panel.

Outcomes

The primary endpoint in this study was the initiation of vasopressor therapy within 24 h after ED arrival. The secondary endpoint in this study was the administration of > 30 mL/kg intravenous fluids within the first 3 h in the ED. Additional endpoints included ICU admission and in‐hospital mortality within 48 h.

Waveform Pre‐Processing, Feature Extraction and Characterization

First, we pre‐processed the continuous raw PPG waveform using MATLAB R2018a (Matlab, the Mathworks, Natick, USA). The initial pre‐processing step involved filtering using a 4th‐order Butterworth band‐pass filter with a frequency range of 0.8 to 20 Hz. The next step involved evaluating the patient's PPG waveform data quality for feature extraction. This assessment was performed by calculating the Signal Quality Index (SQI) [31]. To this end, we first identified the points‐of‐interest (POIs) within the raw PPG waveform, which encompassed the systolic peak, wave onset, dicrotic notch/inflection point, diastolic peak, a‐peak, b‐peak, and e‐peak (Figure 1). Next, a correlation coefficient was computed for each PPG wave by comparing it to an averaged wave template representing the average wave for the specific patient. PPG waves with a correlation coefficient lower than 0.8 were removed from the analysis. This signal was then split into one‐minute windows, resulting in 20 windows per patient in total. For further SQI determination, a one‐minute window was excluded when it contained less than five PPG waves. PPG data was considered low‐quality when it contained less than two 2 min windows in a 20 min time frame. Following this exclusion process, PPG‐derived features were computed per 1 min window, using the identified POIs (Figure 1). After feature extraction, the mean values were calculated over the initial 20 min. For our analysis, we extracted specific features from the PPG waveform itself and its second derivative, the acceleration photoplethysmogram (APG; Table S1 and Figure 1) [27, 28, 32, 33, 34].

Statistical Analysis

Continuous data are presented as medians with interquartile ranges (IQRs), and categorical data as counts and percentages (n, %). Data distribution normality was assessed using the Shapiro–Wilk test. For non‐normally distributed data, the Mann–Whitney U test was used for pairwise comparisons, and the Chi‐square test was used for categorical variables. For comparisons across three groups, the Kruskal‑Wallis test was applied for continuous variables, and the Fisher's exact test was used for categorical variables. PPG‐derived features were scaled and integrated using principal component analysis (PCA) to reduce dimensionality, address collinearity, and capture the majority of variance. Principal components that together explained more than 70% of the total variance were retained for further analysis. A correlation matrix was used to explore relationships between individual PPG‐derived features and principal components. Subsequently, the analysis was then divided into two complementary approaches. First, unsupervised K‐means clustering was applied to the retained principal components to explore whether early septic shock subphenotypes with distinct PPG patterns could be identified. The clinical and hemodynamic characteristics of patients within each cluster were subsequently examined. Second, we conducted supervised multivariable logistic regression models to evaluate the discriminative ability of PPG‐derived components in distinguishing patients who did or did not receive vasopressor therapy within 24 h, as well as other clinical outcomes. Independent variables included PPG‐derived principal components, MAP, and lactate. These models were adjusted for known confounders (e.g., age, gender, pre‐hospital I.V. fluids, cardiovascular comorbidities) and assessed using the Area Under the Receiver Operating Characteristic (AUROC) curve. To assess the diagnostic accuracy of PPG as compared to current parameters, we developed multiple models that integrated MAP, lactate, and PPG in different combinations. The robustness of the PPG model was further evaluated through internal cohort validation (2024). Statistical analyses were performed using RStudio (version 4.3.1), and figures were created using GraphPad Prism 10, Biorender.com, and Adobe Illustrator 2024.

Results

Study Population

We included 325 of 421 eligible patients (77%) from the Acutelines data biobank who were admitted to the ED with a working diagnosis of sepsis requiring hemodynamic resuscitation and available PPG waveforms for analysis (Figure 2). The remaining 96 patients (23%) were excluded for the following 66 patients (16%) were excluded due to a non‐ICU treatment policy, which meant no use of vasopressors, and 30 patients (7%) were excluded due to absent or low‐quality PPG waveforms in the first 20 min. The cohort was 41% female, with a median age of 63, IQR: [50–73]. Relevant comorbidities included 120 patients (37%) with hypertension, 34 (11%) with ischemic heart disease, 41 (13%) with heart failure, and 90 (28%) with diabetes. The median sequential organ failure assessment (SOFA) score was 4 [2, 3, 4, 5, 6, 7]. In terms of hemodynamic management, patients received a median pre‐hospital of 0.5 [0.0, 0.5] L of I.V. fluids, and in ED a median 0.6 [0.1, 1.5] L of I.V. fluids, with 63 (19.4%) receiving more than 30 mL/kg/3 h. Among all patients, 53 (16%) received vasopressor therapy within 24 h, while 272 (84%) did not. Additionally, 82 (25%) patients were admitted to the ICU within 48 h, and 14 (4%) died within 48 h. The diagnosis of an infection was confirmed by a post hoc adjudication 107 patients (33%) had a respiratory focused infection, 48 (15%) had a urinary tract infection, 87 (27%) had other sites of infection, and 83 (26%) had no confirmed infection (Table 1).

Following post hoc infection adjudication, we compared baseline characteristics, vital signs, and PPG‐derived features between patients with confirmed infection and those without (Table S1). The confirmed infection group showed lower pulse width (PW), reflection index (RI), and inflection point area (IPA), indicating increased vasodilation, and a shorter pulse interval (PI), consistent with tachycardia. Lactate was higher in the no‐infection group, possibly due to other acute aetiologies such as diabetic ketoacidosis. Despite fewer I.V. fluids, their outcomes were similar. As clinical decisions rely on infection suspicion at ED arrival, all patients were included in the main analysis to reflect real‐world practice.

Principal Component Analysis and Clustering Reveal Hemodynamic Profiles

After preprocessing, PPG‐derived features were scaled and analyzed using principal component analysis (PCA), with the first three components explaining 80.3% of the variance (PC loadings in Table 2 and Figure 3). PC1 was primarily characterized by PW (−0.487) and RI (−0.451), reflecting arterial compliance. PC2 was mainly characterized by Systolic Peak Amplitude (SPA, 0.566) and Perfusion Index (PPI, 0.578), which relate to cardiac output (CO) and systemic vascular resistance (SVR). PC3 was largely shaped by Diastolic Peak Amplitude (DPA, 0.442) and Delta Time (DT, −0.530), reflecting peripheral vasomotor tone, associated with vasodilation in sepsis. Correlations between PPG‐derived features and principal components were consistent with PC loadings. PC1 correlated strongly with PW (ρ = 0.95), IPA (ρ = 0.80), PI (ρ = 0.72), and RI (ρ = 0.92); PC2 with SPA (ρ = 0.94) and PPI (ρ = 0.96); and PC3 with DPA (ρ = 0.56) and DT (ρ = 0.59; Figure S1).

Unsupervised K‐means clustering of the three principal components identified three distinct clusters, with patient demographics, comorbidities, and clinical characteristics summarized in Table 3. Cluster C exhibited the most severe clinical profile, with lower blood pressure, higher heart rate, elevated respiratory frequency, and lower temperature. These patients also had elevated blood gas lactate, leukocyte, and creatinine levels, as well as the highest clinical severity scores (Table 3). Moreover, patients in Cluster C had the worst outcomes, including the highest rates of vasopressor therapy initiation within 24, 48 ICU admission, and 48 h mortality (Figure 4). In contrast, both Cluster A and Cluster B represented patients with early hemodynamic instability with better clinical outcomes than Cluster C. Despite these differences, the volume of intravenous fluid resuscitation during ED admission did not differ significantly between clusters (Figure 4). In summary, Cluster C appears more severely ill and in greater need of hemodynamic support (i.e., vasopressor therapy, ICU admission) despite similar fluid volumes, suggesting these patients may be experiencing refractory vasodilatory shock rather than hypovolemic shock, with instability driven by factors like vasodilation due to disturbed vasomotor tone in sepsis. A limited subgroup analysis comparing patients with no usable PPG signal to those in Cluster C showed comparable clinical characteristics (Table S2), suggesting that absent or poor‐quality PPG may also indicate underlying hemodynamic compromise.

To further characterize the hemodynamic differences between clusters, we examined the distribution of individual PPG‐derived features, illustrated in Figure 5. Patients in Cluster A had, on average, the lowest IPA, crest time (CT), RI, PW, and APG b/a ratio compared to the other clusters. Cluster B had the highest DPA and broader PI. Cluster C exhibited the lowest SPA, DPA, DT, and PPI. These features in Cluster C are linked to impaired cardiac output, decreased systemic vascular resistance and disturbed peripheral vasomotor tone, indicating a hemodynamic state more suggestive of predominantly vasodilative shock (Figure 5). To complement these findings, Figure 6 displays representative PPG waveforms from patients closest to each cluster centroid, highlighting morphological patterns that correspond to the identified PPG‐derived features, representing distinct hemodynamic profiles.

PCA Regression Model and Discriminative Power

Multivariable logistic regression was performed using PC1–PC3, which together explained 80.3% of the PPG variance. To adjust for confounders, we included age, gender, and cardiovascular comorbidity (mild: hypertension/diabetes mellitus; ischemic heart disease; heart failure), as well as pre‐hospital I.V. fluids.

The PPG‐only model demonstrated moderate discriminative power for identifying patients who received vasopressor therapy within 24 h (AUROC: 0.746), with significant contributions from PC2 and PC3 (p < 0.05). Its performance was comparable to the classic MAP‐only model (AUROC: 0.71; p = 0.496). The MAP + lactate model showed similar performance (AUROC: 0.75), while adding lactate to PPG improved discrimination (AUROC: 0.79). The combined PPG + MAP + lactate model performed best (AUROC: 0.83), with a sensitivity of 67%, specificity of 84%, positive predictive value (PPV) of 96%, and negative predictive value (NPV) of 33%, significantly outperforming both the PPG‐only and MAP + lactate models (p = 0.010), which did not differ from each other (p = 0.824). Performance metrics for all models are detailed in Figure 7.

The combined model also showed moderate discriminative power for secondary endpoints, including patients who required higher‐volume fluid resuscitation (> 30 mL/kg, AUROC: 0.73) and were admitted to the ICU within 48 h (AUROC: 0.77). The discriminative power for in‐hospital mortality was good (AUROC: 0.86). Receiver‐operating characteristic (ROC) curves for secondary endpoints and corresponding performance metrics are presented in Figure S2.

Internal validation with a separate 2024 cohort was also performed, as outlined in Table S3 which includes study population characteristics, and Figure S3, where sensitivity, specificity, PPV, NPV, and accuracy metrics highlight the model's robustness across cohorts. In the internal validation cohort, the combined model again had the highest AUROC but did not significantly outperform the MAP + lactate model, likely due to the smaller sample size and imbalanced outcome distribution (5% vasopressor use, Table S3). Importantly, it did not perform worse, supporting the robustness and added value of PPG‐derived features. Notably, the PPG‐only model performed as well as the MAP + lactate model (Figure S3), reinforcing the potential of PPG to capture peripheral perfusion dynamics in septic shock and serve as a non‐invasive alternative to established clinical parameters.

Discussion

Main Findings

This study investigated whether PPG‐derived features measured during the first 20 min of ED admission could identify early septic shock subphenotypes, and could help identify patients with predominant vasodilatory shock requiring vasopressor therapy within 24 h. First, unsupervised K‐means clustering using PPG components identified three patient subphenotypes with partially overlapping characteristics but distinct clinical and hemodynamic profiles, including differences in age, hypertension, blood pressure, lactate, and creatinine. These differences underscore the ability of PPG‐based clustering to capture clinically meaningful variations among patients. Cluster C comprised patients in septic shock, characterized by worsened vital signs, a greater need for hemodynamic support therapy, ICU admission, and increased mortality. Additionally, patients with absent or poor‐quality PPG signals show characteristics similar to Cluster C, suggesting underlying hemodynamic compromise and supporting prior findings on the prognostic value of absent pulse oximeter signals in prehospital care [35]. In contrast, Clusters A and B consisted of patients with suspected sepsis and early hemodynamic instability, who received similar volumes of I.V. fluids but experienced better outcomes than Cluster C. While these clusters are not sharply defined, they reflect a spectrum of hemodynamic changes in sepsis, from a more normal pattern to a vasodilatory state with or without concurrent hypovolemia, suggesting that non‐invasive PPG could offer valuable insights for tailoring sepsis management based on a patient's specific hemodynamic profile.

Second, the supervised multivariable analysis showed that a discriminative model based solely on PPG‐derived features demonstrated moderate accuracy in identifying patients who required vasopressor therapy within 24 h. However, a model combining PPG with MAP proved to be more accurate in identifying patients who required vasopressors, suggesting that PPG provides additional valuable insights beyond MAP alone. MAP reflects the interplay between cardiac output and systemic vascular resistance, while PPG can detect peripheral pathophysiological changes, such as vasodilation. This suggests that integrating PPG could help recognize patients with early septic shock characterized by a predominant vasodilatory component who might benefit from vasopressor therapy early, thereby allowing for more tailored hemodynamic resuscitation and improving sepsis outcomes.

Pathophysiological Explanation

PPG‐derived features can be grouped via PCA into distinct components that further refine our understanding of the hemodynamic state. For instance, PC1, primarily characterized by PW and RI, is associated with arterial compliance [26, 27]. PC2, encompassing features such as PPI and SPA, reflects peripheral perfusion influenced by both CO and SVR [28, 34, 36]. PC3, reflecting DPA and DT, provides insight into vasomotor tone alterations such as vasodilation, where fluid resuscitation is unlikely to be effective and early vasopressor initiation may be preferable [28, 37, 38]. This physiological relevance underscores the potential of PPG as a clinical decision support tool that continuously and non‐invasively capturing dynamic hemodynamic changes through peripheral perfusion measurements at the fingertip.

Although MAP provides insight into hemodynamic status, it primarily reflects central circulatory function and does not fully capture peripheral changes [39, 40]. Similarly, lactate levels are an important biomarker for hypoperfusion and organ dysfunction in sepsis, but they do not offer a comprehensive view of cardiovascular dynamics [41, 42]. In fact, a recent study showed that continued resuscitation on lactate levels in patients with normal peripheral perfusion might even be harmful [42]. Moreover, their findings suggest that a resuscitation strategy targeting peripheral perfusion status may be superior to one based solely on lactate levels [43]. In this context, PPG provides a continuous, non‐invasive method to assess peripheral perfusion and hemodynamic changes, thereby potentially serving as a valuable tool to guide personalized hemodynamic resuscitation in sepsis. Integrating PPG‐derived components with MAP and lactate may thus offer a more comprehensive hemodynamic profile, potentially improving the ability to recognise patients with a predominant vasodilatory profile requiring earlier vasopressor therapy, and to guide tailored management strategies in sepsis.

Clinical Relevance

This study emphasizes the need for a personalized approach to hemodynamic resuscitation in sepsis. PPG can recognize dynamic hemodynamic changes in sepsis, allowing the identification of hemodynamic profiles and distinguishing between different subphenotypes of septic shock. Sepsis often presents as a combination of hypovolaemic and distributive shock, requiring a balance of therapeutic vasopressor therapy for the vasodilatory component and volume resuscitation to correct hypovolaemia [24]. While a personalized approach is essential, traditional measures such as MAP and lactate have limitations in assessing the full spectrum of circulatory dysfunction [20, 42, 44]. PPG, however, offers additional real‐time, non‐invasive insights into cardiovascular status in sepsis [45, 46]. An additional advantage is its ability to provide continuous monitoring through wearable devices, which can be used not only in the ED but also in prehospital care, general practice, and home care settings [47]. This capability supports PPG as a promising clinical decision support tool to facilitate the early identification of predominant shock type and guide clinical decision‐making, ensuring the selection of the most appropriate therapeutic intervention for the individual, and optimizing clinical outcomes in sepsis management.

Future Implications

PPG offers a non‐invasive, continuous approach to monitor cardiovascular dynamics, and holds promise as a clinical decision support tool for personalized hemodynamic resuscitation in sepsis. This proof‐of‐principle study establishes an important foundation but further steps are necessary to advance clinical implementation. Future research should first perform external validation to confirm that PPG reliably captures cardiovascular characteristics across diverse patient populations [48, 49]. Second, model development should focus on creating clinically interpretable algorithms by selecting important features, defining actionable thresholds, and exploring computational methods suitable for real‐time decision support [49, 50, 51]. Third, clinical implementation requires optimizing and automating the signal processing pipeline for near real‐time application at bedside, potentially extending to wearable devices [51, 52, 53]. Fourth, clinical trial enrichment should evaluate the role of PPG in predictive and prognostic enrichment [54, 55], particularly to guide early vasopressor initiation in patients with predominant vasodilatory septic shock. Finally, broadened application should investigate the additional value of PPG in assessing fluid responsiveness and monitoring treatment effects, as previous studies have linked the perfusion index to fluid responsiveness in ICU patients with septic shock [56, 57]. In summary, future work should focus on improving the technique and methodology, performing robust external validation, and developing practical real‐time bedside tools with actionable thresholds for integration into sepsis resuscitation protocols.

Limitations

Several limitations should be acknowledged in this study. Firstly, it was conducted in a single tertiary care centre. However, the large rural/urban catchment area ensures that the case mix is representative of both general and academic teaching hospitals. As an observational study, missing data posed a challenge. Nonetheless, essential information was manually verified to minimize its impact. While missing data is an inherent limitation in real‐world studies, it also increases the likelihood of successful implementation in clinical practice. Additionally, to minimize technical artifacts, a strict SQI algorithm (0.8) was applied, resulting in the exclusion of some data from the analysis. This approach ensured the inclusion of high‐quality data, even for patients with low peripheral perfusion due to septic shock. Only 7.1% of the patients were excluded due to absent or poor quality, in the first 20 min, and this quality control procedure is not expected to introduce bias. Notably, a subgroup analysis showed that patients without usable PPG signals had clinical characteristics similar to those in the septic shock subphenotype, indicating that poor‐quality signals may also reflect hemodynamic compromise.

Conclusion

In conclusion, this proof‐of‐principle study demonstrated that early, non‐invasive photoplethysmography (PPG) measurement during emergency department triage can identify septic shock subphenotypes and moderately distinguish patients who received subsequent vasopressor therapy, likely due to a vasodilatory profile, thus providing additional value beyond classical parameters such as mean arterial pressure and lactate. By early identification of vasodilatory shock, with or without a hypovolemic component, through the continuous assessment of peripheral perfusion and hemodynamic changes at the fingertip, PPG may help identify patients who might benefit from early initiation of vasopressor therapy as a part of a personalized hemodynamic resuscitation strategy. This approach could support clinical decision making by carefully balancing the choice between higher‐volume fluid resuscitation and early initiation of vasopressor therapy, ultimately optimizing sepsis management and potentially improving patient outcomes.

Author Contributions

Sanne Ter Horst: conceptualization, methodology, software, formal analysis, investigation, writing ‑ original draft, visualization, project administration. Anna D. Schoonhoven: methodology, software, investigation, writing ‑ review and editing. Raymond J. van Wijk: methodology, software, investigation, writing ‑ review and editing. Rick Weitering: investigation, data curation, writing ‑ review and editing. Sanne W. van Loon: investigation, data curation, writing ‑ review and editing. Jan C. ter Maaten: methodology, conceptualization, writing ‑ review and editing, supervision. Hjalmar R. Bouma: conceptualization, methodology, validation, writing ‑ review and editing, supervision, project administration.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information