Visual scoring of photoplethysmography morphology for risk stratification in hemorrhagic shock: a prospective observational study

Highlights

• •Visual PPG scoring showed good prediction of mortality and ICU admission
• •PPG morphology may aid early bedside risk stratification in hemorrhage
• •Inter-rater agreement and repeatability for visual PPG scoring was good
• •Visual PPG scoring outperformed conventional hemodynamic indicators

Abstract

Purpose

Early recognition of high-risk patients with hemorrhage remains challenging, as conventional indicators may lack sensitivity during the compensated phase of blood loss. The study evaluated whether visual scoring of photoplethysmography (PPG) morphology could support bedside risk stratification during acute hemorrhage.

Methods

In this prospective observational study conducted in the emergency department of West China Hospital between August and October 2025, 500 adults with hemorrhage were enrolled, of whom 461 were included in the final analysis. Three trained physicians independently scored PPG morphology using an ordinal 0–100 visual scale in 10-point increments. The final score was determined through majority voting. The primary outcome was 28-day all-cause mortality, while the secondary outcome was intensive care unit (ICU) admission. Moreover, lactate-related risk stratification was explored separately.

Results

Inter-rater agreement was good to excellent based on pairwise linear weighted kappa (0.74, 0.99, and 0.75), whereas overall agreement was moderate based on Fleiss’ kappa (0.58). The visual PPG score demonstrated good discrimination for both 28-day mortality (area under the curve [AUC] 0.81) and ICU admission (AUC 0.88), with numerically superior performance to the shock index, mean arterial pressure, lactate, and base excess. Adding the visual PPG score to routine clinical indicators increased the AUC for ICU admission from 0.82 to 0.90.

Conclusion

Visual PPG scoring may support early bedside risk stratification in hemorrhagic patients and warrants external validation.

Introduction

Early risk stratification in acute hemorrhage remains difficult. Current assessment relies largely on estimated blood loss and changes in blood pressure, which are often insufficient for the timely identification of high-risk patients [1,2]. This is largely attributed to the substantial physiological compensation occurring early following hemorrhage [3]. Thus, prognosis depends not only on the extent of blood loss, but also on the capacity to maintain circulatory stability, referred to as hemorrhagic compensatory capacity [3,4]. Conventional hemodynamic variables such as blood pressure and heart rate reflect the overall circulatory state at a single time point and may inadequately capture the progressive transition from compensation to decompensation [2,4]. More sensitive indicators of compensatory status are therefore needed.

Photoplethysmography (PPG) is a noninvasive signal that reflects changes in peripheral blood volume [5]. Beyond pulse detection alone, PPG morphology carries information on peripheral perfusion, vascular tone, and pulse wave reflection [[5], [6], [7]]. Studies of compensatory reserve and automated waveform analysis have demonstrated that features such as amplitude, pulse contour, and dicrotic morphology change with worsening hemodynamic status. However, most existing evidence is derived from simulated hemorrhage, laboratory models, or computationally intensive frameworks, which may not fully represent the physiological complexity of real-world hemorrhage [[8], [9], [10]]. In clinical practice, pain, inflammation, and neuroendocrine activation may influence vascular tone, peripheral perfusion, and waveform morphology [7,10].

In this study, we evaluated a simple morphology-based visual scoring system for PPG in patients with hemorrhage. We hypothesized that visually recognizable waveform patterns may provide clinically useful bedside risk stratification beyond routine shock assessment.

Methods

Study design and setting

This prospective observational study was conducted at the Emergency Department of West China Hospital, Sichuan University, between August 1 and October 31, 2025. The study was approved by the Biomedical Ethics Review Committee of West China Hospital (Approval No. 2025 Annual Review [1245]) and registered in the Clinical Trial Registry (ChiCTR2500107411). All data were anonymized before analysis.

Participants

Adults aged 18–80 years presenting with suspected or clinically apparent acute hemorrhage were screened for inclusion. Patients were excluded if they were pregnant, had acquired immunodeficiency syndrome, or if the final diagnostic evaluation failed to confirm bleeding or identified a condition unrelated to hemorrhagic shock. Patients with incomplete data or loss to follow-up were excluded from the final analysis. Of the 500 enrolled patients, 461 were included in the final cohort.

Data collection and PPG acquisition

At presentation, demographic characteristics, conventional vital signs, and laboratory variables were collected, including heart rate, mean arterial pressure, respiratory parameters, hemoglobin, blood urea nitrogen (BUN), creatinine, albumin, glucose, bilirubin, triglycerides, total cholesterol, lactate, and base excess, as available. A 3-minute PPG recording was obtained during triage along with routine blood gas and laboratory testing. PPG signals were acquired using a pulse oximeter finger probe (512F) connected to an ePM 10M patient monitor (Mindray, Shenzhen, China) used for emergency monitoring. Clinical management followed standard emergency practice and was not influenced by the study assessments.

Visual scoring of PPG morphology

PPG waveforms were assessed using a morphology-based visual scoring system. The three raters were physicians with relevant clinical expertise. Before formal scoring, they underwent a standardized 2 -h training session on waveform interpretation based on predefined visual scoring criteria. The scoring instrument was developed de novo for this study from the physiological principles described above; a complete descriptive rubric for all 11 ordinal categories is provided in Supplementary Table S2. After training, the raters independently evaluated each waveform. Each rater reviewed the entire 3-minute recording and selected a representative segment (approximately 10–20 seconds) without artifacts based on preset quality standards: a stable baseline, minimal motion artifacts, consistent pulse morphology, and a clearly distinguishable outline. Segments of poor quality were excluded. The waveform was then assigned to one of 11 ordinal categories ranging from 0 to 100 in 10-point increments based on amplitude, waveform width, dicrotic morphology, and the steepness of the systolic upstroke. The raters were independent of the clinical data collection team. The final score was determined by majority voting; if all three raters assigned different categories, the median category was selected. When the four morphological features were discordant, raters prioritized the overall severity pattern most consistent with hypoperfusion. Higher scores corresponded to greater waveform amplitude, clearer contour definition, and more recognizable dicrotic morphology, whereas lower scores corresponded to attenuated amplitude, waveform flattening, a blunted upstroke, and loss of dicrotic features (Fig. 1).

Fig. 1.

Representative PPG waveforms corresponding to different visual score categories. The complete descriptive scoring rubric for all 11 ordinal categories is provided in Supplementary Table S2.

Blinding and outcomes

The three raters were blinded to lactate levels, intensive care unit (ICU) admission, and mortality outcomes at the time of waveform scoring. Treating clinicians did not have access to the visual PPG scores, and outcome adjudication was conducted independently. The primary outcome was 28-day all-cause mortality; meanwhile, the secondary outcome was ICU admission. Exploratory analyses assessed lactate-related risk stratification using lactate thresholds of ≥2 mmol/L and ≥4 mmol/L, consistent with prior studies in trauma and hemorrhagic shock [11,12]. Lactate served two distinct roles: as a comparator predictor in the primary mortality and ICU-admission analyses, and as a binary outcome (lactate ≥4 mmol/L) in the exploratory lactate-related analysis. In addition, we performed an exploratory analysis of subsequent vasopressor-requiring shock (ICU admission with documented vasopressor use). Additional subgroup analyses excluding patients with tumors were also performed and are reported as post-hoc exploratory comparisons alongside the full-cohort primary analyses.

Sample size estimation

Given the scarcity of comparable prior studies, traditional sample-size calculation methods based on effect size were deemed not feasible. Instead, a clinically oriented statistical approach was adopted as an orienting feasibility estimate rather than a formal a priori power calculation. We referenced a similar study evaluating shock prediction models in pediatric patients with polytrauma (n = 325, area under the curve [AUC] = 0.96) [13]. We planned to enroll 500 patients with hemorrhagic conditions over a fixed 3-month study period (August to October 2025), with an expected shock incidence of 16% (approximately 80 positive events) based on the design-stage assumption in the cited pediatric report [14]. This target provided a pragmatic frame for recruitment. Post-hoc power based on the observed event rates in the final cohort (55 deaths and 101 ICU admissions among 461 included patients) was 87% to detect an AUC of at least 0.70 for mortality and more than 99% to detect an AUC of at least 0.70 for ICU admission (alpha = 0.05, two-sided). The observed rate of subsequent vasopressor-requiring shock was lower than the 16% design-stage assumption (2.6% vs. 16%), reflecting our narrower definition based on documented vasopressor use; the shock-discrimination analysis is therefore exploratory.

Statistical analysis

Statistical analyses were performed using Microsoft Excel 2019 and SPSS version 26.0. Continuous variables were tested for normality using the Shapiro–Wilk test and are presented as mean ± standard deviation or median (interquartile range), as appropriate. Moreover, between-group comparisons were performed using the independent-samples t test or the Mann–Whitney U test. Categorical variables are presented as counts and percentages and were compared using the chi-squared test or Fisher’s exact test, as appropriate. The visual PPG score was analyzed as an ordered numerical variable. Tied scores were retained without arbitrary separation; average ranks were assigned in rank-based comparisons, and tied predictor values were handled using the standard empirical procedure in ROC analyses.

Inter-rater agreement was assessed using pairwise linear weighted kappa coefficients, and overall agreement among the three raters was evaluated using Fleiss’ kappa.

Variables associated with ICU admission in univariable analysis (p < 0.05) were entered into a multivariable logistic regression model using the forward likelihood ratio method. The resulting routine clinical model included shock index (SI), BUN, lactate, and albumin and is reported in the Supplementary Material. To assess the robustness of variable selection and protect against overfitting, we additionally performed LASSO logistic regression with 10-fold cross-validation, a smaller pre-specified clinical model (age, SI, lactate, and base excess), and 1000-bootstrap internal validation of the combined PPG-plus-routine model with optimism correction. The results are summarized in Supplementary Table S2.

Results are reported as odds ratios with 95% confidence intervals (CIs). Predictive performance was assessed using receiver operating characteristic (ROC) curves and AUC. Pairwise comparisons of AUCs were performed using the DeLong test. Optimal cutoff values were determined independently for each outcome using the Youden index. A two-sided p-value < 0.05 was considered statistically significant.

Results

Study population

Between August and October 2025, 500 patients with confirmed suspected or clinically apparent acute hemorrhage were enrolled. After excluding 35 patients whose final diagnoses did not confirm bleeding or were unrelated to hemorrhagic shock and four patients who were lost to follow-up, 461 patients remained in the final analysis. Gastrointestinal bleeding accounted for 408/461 cases (88.50%), trauma for 42/461 (9.11%), and other causes of hemorrhage for 11/461 (2.39%). The mean age was 56.83 years, while 323 patients (70.07%) were male. During the 28-day follow-up period, 55 patients (11.93%) died, and 101 (21.91%) were admitted to the ICU (baseline characteristics are presented in Table 1).

Table 1.

Baseline characteristics.

Baseline Characteristics	Overall (n = 461)	Survivors (n = 406)	Non-survivor (n = 55)	P value
Age (years)
18−60	243(52.71%)	215(52.96%)	28(50.90%)	0.781
>60	218(47.29%)	191(47.04%)	27(49.10%)
Time from onset to admission (hours)
≤24 h	166(36.01%)	147(32.20%)	19(34.55%)	0.812
>24 h	295(63.99%)	259(63.79%)	36(65.45%)
Gender (n, %)
Male	323(70.07%)	284(69.95%)	39(70.91%)	0.886
Female	138(29.93%)	122(30.05%)	16(29.09%)
Shock (n, %)
Yes (SI ≥ 1.0)	124 (26.90%)	92 (22.66%)	32 (58.18%)	<0.001
No (SI<1.0)	337 (73.10%)	314 (77.34%)	23 (41.82%)
ICU admission (n, %)
Yes	101 (21.91%)	61 (15.02%)	40 (72.73%)	<0.001
No	360 (78.09%)	345 (84.98%)	15 (27.27%)
Systolic BP (mmHg)	115.12 ± 22.36	116.46 ± 22.02	105.22 ± 22.57	<0.001
Diastolic BP (mmHg)	66.94 ± 13.90	67.79 ± 13.60	60.73 ± 14.58	<0.001
Mean arterial pressure (mmHg)	83.00 ± 15.39	84.01 ± 15.00	75.56 ± 16.32	<0.001
Heart rate (beats/min)	94.91 ± 20.68	93.59 ± 20.54	104.67 ± 19.15	<0.001
Shock index	0.82 (0.65–1.02)	0.80 (0.64–0.97)	1.05 (0.75–1.23)	<0.001
Lactate (mmol/L)	1.70 (1.20–2.50)	1.60 (1.20–2.30)	2.60 (1.65–5.10)	<0.001
Base excess (mmol/L)	3.67 ± 4.86	3.36 ± 4.40	6.00 ± 7.10	0.010
Hemoglobin (g/L)	73.00 (58.00–98.00)	75.50 (60.00–99.75)	62.00 (52.50–73.50)	<0.001
BUN (mmol/L)	7.40 (5.10–11.30)	7.20 (4.90–10.90)	8.60 (6.65–14.35)	0.001
Albumin (g/L)	32.20 ± 6.14	32.85 ± 6.03	27.42 ± 4.69	<0.001
Hematocrit	0.23 (0.19–0.31)	0.24 (0.19–0.31)	0.21 (0.17–0.23)	<0.001
Creatinine (μmol/L)	75.00 (58.00–100.00)	74.00 (57.00–95.00)	102.00 (64.50–176.50)	0.002
Glucose (mmol/L)	7.32 (5.83–9.92)	7.16 (5.79–9.46)	9.62 (6.88–13.12)	<0.001
Total bilirubin (μmol/L)	12.80 (7.40–23.30)	12.80 (7.35–20.65)	15.00 (7.67–37.80)	0.209
Direct bilirubin (μmol/L)	4.70 (2.70–10.10)	4.50 (2.60–8.90)	7.10 (3.32–27.43)	0.003
Triglycerides (mmol/L)	1.13 (0.80–1.55)	1.14 (0.80–1.56)	1.12 (0.79–1.46)	0.635
Cholesterol (mmol/L)	2.79 (2.22–3.52)	2.91 (2.31–3.60)	2.20 (1.64–2.85)	<0.001
HDL-C (mmol/L)	0.77 (0.57–1.00)	0.80 (0.61–1.02)	0.55 (0.43–0.72)	<0.001
LDL-C (mmol/L)	1.42 (0.92–1.99)	1.46 (1.04–2.06)	0.85 (0.58–1.45)	<0.001
INR	1.13 (1.01–1.31)	1.12 (1.01–1.27)	1.33 (1.19–1.67)	<0.001
Respiratory rate (breaths/min)	20.00 (20.00–22.00)	20.00 (19.75–21.00)	20.00 (20.00–22.00)	0.273
Classification of Bleeding Etiologies (n, %)
Gastrointestinal Bleeding	408(88.50%)	364(89.66%)	44(80%)	0.063
Traumatic bleeding	30 (6.51%)	25 (6.16%)	5 (9.09%)
Others	23 (4.99%)	17 (4.19%)	6 (10.91%)
Comorbidities (n, %)
Liver cirrhosis	129 (27.98%)	117 (28.82%)	12 (21.82%)	0.278
Hypertension	110 (23.86%)	99 (24.38%)	11 (20.00%)	0.474
Malignancy	129 (27.98%)	107 (26.35%)	22 (40.00%)	0.034
Diabetes	57 (12.36%)	48 (11.82%)	9 (16.36%)	0.337
Hepatitis B	33 (7.16%)	29 (7.14%)	4 (7.27%)	1.000
Renal insufficiency/failure	27 (5.86%)	17 (4.19%)	10 (18.18%)	<0.001
Coronary heart disease	26 (5.64%)	26 (6.40%)	0 (0.00%)	0.058

Inter-rater reliability

Weighted kappa demonstrated good agreement between raters 1 and 2 (0.74) and between raters 2 and 3 (0.75), and excellent agreement between raters 1 and 3 (0.99). Overall agreement among the three raters was moderate, with a Fleiss’ kappa of 0.58.

Predictive performance for mortality and ICU admission

The visual PPG score exhibited good discrimination for 28-day mortality (AUC 0.81; 95% CI, 0.76–0.87) and had numerically higher AUC values than the shock index [SI] (0.69), mean arterial pressure (0.68), lactate (0.71), and base excess (0.64). After excluding patients with tumors, the AUC increased to 0.85. At a visual PPG score cutoff of 50, sensitivity and specificity for mortality were 84.91% and 63.40%, respectively (Fig. 2).

Fig. 2.

Visual PPG score for estimating 28-day mortality: (A) entire cohort; (B) subgroup excluding patients with tumors. PPG, visual photoplethysmography score; SI, shock index; MAP, mean arterial pressure; LAC, lactate; BE, base excess.

The visual PPG score exhibited good discrimination for ICU admission (AUC 0.88; 95% CI, 0.84–0.92) and had numerically higher AUC values than the SI (0.73), mean arterial pressure (0.64), lactate (0.71), and base excess (0.64). After excluding patients with tumors, the AUC increased to 0.89. At a visual PPG score cutoff of 50, sensitivity and specificity for ICU admission were 88.12% and 70.36%, respectively (Fig. 3).

Fig. 3.

Visual PPG score for estimating ICU admission: (A) entire cohort; (B) subgroup excluding patients with tumors. PPG, visual photoplethysmography score; SI, shock index; MAP, mean arterial pressure; LAC, lactate; BE, base excess.

Incremental value beyond routine indicators

In the multivariable analysis for ICU admission, SI, BUN, lactate, and albumin remained significant predictors. The routine clinical model based on these indicators is demonstrated in the Supplementary Material and was used as the reference model in the ROC analysis. The visual PPG score (AUC 0.88) revealed significantly better discrimination than that observed with the routine clinical model (AUC 0.82; DeLong test, p = 0.01). Adding the PPG score to the routine clinical model increased the AUC to 0.90, which was significantly higher than that of the routine clinical model alone (p < 0.0001) but not significantly different from the PPG score alone (p = 0.12). These findings were supported by robustness analyses (Supplementary Table S2).

Exploratory lactate-related analysis

In this exploratory analysis, lactate ≥4 mmol/L was the binary outcome being predicted by the visual PPG score, not a comparator predictor. Using a lactate threshold of ≥4 mmol/L, the visual PPG score demonstrated an AUC of 0.85, higher than that of the SI (0.74) and mean arterial pressure (0.65) (Fig. 4A). After excluding patients with tumors, the AUC increased to 0.89 (Fig. 4B). At a visual PPG score cutoff of 50, sensitivity and specificity were 87.76% and 63.11%, respectively

Fig. 4.

Visual PPG score for lactate estimation: (A) entire cohort; (B) subgroup excluding patients with tumors. PPG, visual photoplethysmography score; SI, shock index; MAP, mean arterial pressure; LAC, lactate; BE, base excess.

High-risk subgroup

Among the 24 patients with a visual PPG score below 20, 28-day mortality was 75.0% and ICU admission was 90.5%, indicating a particularly high-risk subgroup.

Discussion

In this prospective observational study, the visual PPG score exhibited good discrimination for 28-day mortality and ICU admission and improved predictive performance when added to routine clinical indicators. Combining the PPG score with traditional clinical indicators significantly improved ICU admission prediction compared with routine indicators alone; however, the combined model did not significantly outperform the PPG score alone (p = 0.12). This pattern is consistent with two interpretations: the visual PPG score may add incremental information to routine assessment, or it may capture most of the prognostically relevant information in this cohort with limited incremental gain from adding routine indicators. The non-significant comparison does not establish equivalence and may reflect limited power to detect a smaller incremental benefit.

A major advantage of pulse waveform assessment is its ability to reflect microcirculatory changes more directly than traditional methods based on systemic hemodynamic parameters. As emphasized in microcirculation research, microcirculation represents the critical interface between systemic hemodynamics and tissue oxygenation. In hemorrhagic shock, microcirculatory dysfunction often precedes measurable changes in systemic vital signs [15]. Traditional indicators, including central venous pressure, cardiac output, heart rate, and blood pressure, primarily reflect systemic circulation. In the early phase after bleeding onset, systemic hemodynamic stability may be temporarily maintained through compensatory vasoconstriction, thereby masking underlying microcirculatory injury. This discrepancy is referred to as hemodynamic separation [16]. The clinical definition of hemorrhagic shock itself remains heterogeneous, and traditional vital signs have limited sensitivity during compensated blood loss [26]. In contrast, pulse wave signals arise from changes in peripheral blood volume and hemoglobin oxygenation and can therefore reflect microcirculatory perfusion, vascular reactivity, and the balance between tissue oxygen supply and demand more directly [17]. This advantage likely reflects the limitations of traditional indicators: in the early to moderate stages of bleeding, these variables are often masked by physiological compensation and may fail to capture subtle but clinically important hemodynamic changes. Conversely, PPG waveforms reflect alterations in arterial flow patterns during bleeding, including reduced amplitude and blunting of the dicrotic notch, enabling more sensitive detection of early shock states. In this study, descriptors such as “reduced amplitude” refer to cross-sectional comparisons within the cohort, not to changes from an individual patient baseline. Pairwise weighted kappa values indicated good to excellent inter-rater agreement, supporting reproducibility of the ordinal scoring framework, although lower Fleiss’ kappa suggests that exact agreement across all three raters was less consistent.

Reduced waveform amplitude represents diminished stroke volume and peripheral perfusion, whereas waveform flattening, dicrotic blunting, and altered upstroke configuration reflect increased vasoconstriction, impaired arterial filling, and progression toward hemodynamic decompensation [[18], [19], [20], [21], [22], [23]]. The visual score may therefore capture integrated morphological manifestations of compensatory status rather than relying on a single threshold-based variable [[19], [20], [21], [22],24]. The performance of the modality under real-world clinical conditions, where compensatory reserve is influenced by factors beyond blood loss alone, further supports its relevance in heterogeneous hemorrhagic patients [20,24].

The visual PPG score is best viewed as complementary rather than substitutive. Lactate remains a well-established marker of tissue hypoperfusion [11,12,19,25], but it requires invasive sampling and may be less readily available during early triage or in prehospital settings. By contrast, the visual PPG score is derived from a noninvasive bedside signal and can be obtained rapidly in clinical practice. The improved performance of the combined model over routine indicators alone indicates that the visual PPG score adds clinically relevant information to routine shock assessment, particularly when rapid bedside evaluation is needed [24]. However, because the combined model did not significantly outperform the PPG score alone, the score may also be used as a standalone bedside risk stratification tool pending external validation.

Limitations

Several limitations of this study should be acknowledged. This was a single-center study, and most patients had gastrointestinal bleeding, potentially limiting the generalizability of our findings to trauma-predominant populations. Although pairwise weighted kappa values were favorable, visual scoring remains observer-dependent, and the lower Fleiss’ kappa indicates that exact agreement across all three raters was less consistent than pairwise ordinal agreement. In addition, the study did not assess whether real-time use of the visual PPG score influences clinical decision-making or patient outcomes. A single triage recording cannot distinguish acute hemorrhage-related waveform changes from inter-individual variation due to age, vascular stiffness, comorbidities, and medications; therefore, the score should be interpreted as a cross-sectional risk indicator rather than proof of deterioration from an individual baseline. The sensitivity analysis excluding patients with tumors was post-hoc and exploratory, not pre-specified in the protocol; full-cohort results are presented as the primary findings. The exploratory analysis of subsequent vasopressor-requiring shock was limited by a low event rate (12 events, 2.6%) and should be interpreted with caution. Multicenter validation is therefore needed.

Conclusion

The visual PPG score may support early bedside risk stratification in hemorrhagic patients and provide clinically relevant information beyond routine shock assessment. It may be used as a standalone bedside tool or as part of a broader clinical assessment, pending external validation.

CRediT authorship contribution statement

*Peng Deng and Xuanyi Xiang contributed equally to this work.

• •Peng Deng: Conceptualization, Methodology, Investigation, Validation, Writing (original draft).
• •Xuanyi Xiang: Investigation, Data curation, Formal analysis, Visualization, Writing (original draft).
• •Yanyan Yu: Investigation.
• •Haifang Yu: Resources, Supervision, Writing (review and editing).
• •Yu Cao: Resources, Supervision, Funding acquisition, Writing (review and editing).
• •Kelin Jia: Resources.

Funding sources

This study was supported by grant [GYYX24017] from the 1·3·5 Project of Center for High Altitude Medicine, West China Hospital, Sichuan University.

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declaration of competing interest

The authors declare that they have no competing interests.

Appendix A. Supplementary data

The following are Supplementary data to this article: