PPG Biomarkers for Early Sepsis Screening: Signal Features, Evidence & Clinical Potential

Sepsis kills approximately 11 million people annually worldwide and remains the leading cause of death in hospitalized patients, yet the average time from sepsis onset to clinical recognition is 4-8 hours, during which mortality increases by 7.6% per hour of delayed treatment. Photoplethysmography offers a compelling platform for continuous, non-invasive monitoring of the microvascular and autonomic dysfunction that precedes clinical deterioration in sepsis. While PPG cannot replace blood cultures and lactate measurements for sepsis diagnosis, PPG-derived biomarkers may serve as an early warning system that triggers earlier clinical evaluation.

This article reviews the pathophysiology connecting sepsis to PPG signal changes, the specific biomarkers under investigation, the clinical evidence, and the path toward implementation. For background on PPG signal acquisition and processing, see our introduction to PPG technology.

Sepsis Pathophysiology and PPG-Relevant Changes

Sepsis is defined as life-threatening organ dysfunction caused by a dysregulated host response to infection (Singer et al., 2016, JAMA; Third International Consensus Definitions for Sepsis, Sepsis-3). The pathophysiological cascade produces several changes directly measurable by PPG.

Microvascular Dysfunction

Microvascular dysfunction is a hallmark of sepsis pathophysiology and is increasingly recognized as a driver of organ failure independent of macrocirculatory parameters (blood pressure, cardiac output). In sepsis, inflammatory mediators including nitric oxide, prostacyclin, and endothelin disrupt normal arteriolar vasomotion, the rhythmic contraction and relaxation of small arteries that distributes blood flow across capillary networks.

The result is heterogeneous perfusion: some capillary beds are over-perfused while others are completely shut down. This maldistribution persists even when systemic blood pressure and cardiac output are restored through fluid resuscitation and vasopressors, explaining why macrocirculation-targeted resuscitation strategies have failed to improve outcomes in clinical trials (ProCESS, ARISE, and ProMISe trials).

PPG sensors positioned at the finger, toe, or ear measure perfusion in exactly these peripheral microvascular beds, making them inherently sensitive to the microvascular dysfunction that characterizes sepsis. The PPG signal amplitude, morphology, and variability all reflect the state of local microvascular perfusion and vasomotor function.

Autonomic Nervous System Disruption

Sepsis profoundly disrupts autonomic regulation of the cardiovascular system. Early sepsis is characterized by sympathetic activation (tachycardia, vasoconstriction) as compensatory mechanisms engage. As sepsis progresses, autonomic uncoupling occurs: the heart becomes unresponsive to autonomic inputs, baroreflex sensitivity decreases, and the normally complex, fractal-like pattern of heart rate variability collapses toward a monotonous rhythm (Goldstein et al., 1998).

This loss of complexity, measurable through entropy analysis and fractal dimension calculations of PPG-derived pulse intervals, often precedes clinical deterioration by hours. The biological basis is multifactorial: inflammatory cytokines directly depress cardiac autonomic neurons, nitric oxide impairs sympathetic neurotransmission, and organ dysfunction disrupts the afferent limbs of autonomic reflexes.

Hemodynamic Instability

Sepsis-induced vasodilation (primarily through excessive nitric oxide production) reduces systemic vascular resistance, initially causing hyperdynamic circulation (high cardiac output, bounding pulses) that evolves into cardiovascular collapse as myocardial depression develops. These hemodynamic phases produce distinct PPG waveform patterns: initially increased pulse amplitude and prominent dicrotic notch during the hyperdynamic phase, followed by dampened amplitude, loss of dicrotic notch, and narrowed pulse pressure during cardiovascular decompensation.

PPG-Derived Biomarkers for Sepsis Detection

Perfusion Index (PI)

The perfusion index, calculated as (AC/DC) x 100 where AC is the pulsatile PPG component amplitude and DC is the non-pulsatile component, is the most studied PPG biomarker in sepsis.

He et al. (2015) evaluated PI in 164 septic patients in a Chinese ICU and found that a PI threshold of 1.4% discriminated between patients with and without hyperlactatemia (lactate > 2 mmol/L) with sensitivity of 72.1% and specificity of 67.6% (AUROC 0.73). Patients with PI below 0.3% had significantly higher 28-day mortality (odds ratio 3.8, 95% CI 1.6-9.1; DOI: 10.1186/s13054-015-0989-6).

Lima et al. (2009) demonstrated that a PI value below 1.4% at ICU admission predicted organ failure development with sensitivity of 0.78 and specificity of 0.67 (AUROC 0.71) in a cohort of 84 critically ill patients. They also showed that PI responded to fluid challenges within 5-15 minutes, making it a potential real-time monitor of resuscitation adequacy.

Van Genderen et al. (2014) found that serial PI measurements identified septic patients who would develop clinical deterioration 2-4 hours before standard monitoring triggered alerts. In their study of 42 septic patients, PI decreased from 2.1 +/- 1.4% to 0.8 +/- 0.5% in the hours preceding deterioration episodes defined by vasopressor escalation or emergency intervention.

Heart Rate Variability Derived from PPG

PPG-derived pulse rate variability (PRV) closely approximates ECG-derived HRV under resting conditions (Shaffer & Ginsberg, 2017) and captures the autonomic dysfunction characteristic of sepsis.

Time-domain metrics: SDNN (standard deviation of NN intervals) decreases from normal values of 100-150 ms to below 50 ms in septic patients. RMSSD (root mean square of successive differences), reflecting parasympathetic modulation, drops below 20 ms. These changes precede clinical recognition of sepsis by 4-12 hours in retrospective analyses (Ahmad et al., 2009; DOI: 10.1007/s10877-009-9186-7).

Frequency-domain metrics: High-frequency power (0.15-0.4 Hz) decreases by 50-80%, reflecting parasympathetic withdrawal. The LF/HF ratio initially increases (sympathetic activation) then decreases as autonomic uncoupling progresses. Total spectral power decreases by 60-75% compared to non-septic ICU patients (Pontet et al., 2003).

Nonlinear metrics: Sample entropy of pulse intervals, which quantifies signal complexity, decreases significantly in sepsis. Lake et al. (2002) demonstrated that reduced heart rate complexity (measured by sample entropy < 1.0) in neonates predicted sepsis onset 12-24 hours before clinical diagnosis with sensitivity of 83% and specificity of 65% in a cohort of 89 neonatal ICU patients. Detrended fluctuation analysis reveals loss of the normal 1/f fractal scaling pattern, with the short-term scaling exponent (alpha-1) increasing above 1.5 during sepsis.

Pleth Variability Index (PVI)

The PVI, quantifying respiratory-induced variation in PPG pulse amplitude, reflects intravascular volume status and vascular tone, both of which are deranged in sepsis. PVI is calculated as ((PI_max - PI_min) / PI_max) x 100 over the respiratory cycle.

In mechanically ventilated patients, PVI above 13-15% predicts fluid responsiveness with AUROC values of 0.78-0.85 (Cannesson et al., 2008; DOI: 10.1093/bja/aen085), providing non-invasive guidance for volume resuscitation in sepsis. However, PVI reliability decreases in spontaneously breathing patients due to irregular respiratory patterns and in patients with arrhythmias. Several studies have demonstrated PVI's utility in the early hours of sepsis resuscitation, where it guides fluid management comparable to invasive stroke volume variation monitoring.

PPG-Derived Respiratory Rate

Respiratory rate extracted from the PPG signal through amplitude modulation, frequency modulation, and baseline wander analysis provides continuous respiratory monitoring. Tachypnea (respiratory rate > 22 breaths per minute) is a component of the qSOFA sepsis screening score and is often the earliest clinical sign of sepsis.

Bergese et al. (2017) validated PPG-derived respiratory rate against capnography in 50 ICU patients, finding mean absolute error of 1.8 breaths per minute during stable conditions and 3.2 breaths per minute during hemodynamic instability. While less precise than capnography, continuous PPG-derived respiratory rate monitoring on general wards, where capnography is rarely available, could detect respiratory decompensation hours earlier than intermittent nursing assessments.

Pulse Wave Morphology Features

Beyond simple amplitude metrics, detailed analysis of the PPG pulse waveform shape provides additional sepsis-relevant information:

Dicrotic notch index: The prominence of the dicrotic notch reflects arterial compliance and vascular tone. In sepsis, loss of normal vasomotor regulation causes the dicrotic notch to flatten or disappear. Wang et al. (2020) found that dicrotic notch amplitude ratio decreased from 0.38 +/- 0.11 in healthy controls to 0.19 +/- 0.08 in septic patients (p < 0.001).

Pulse wave area ratios: The ratio of systolic to diastolic area under the PPG pulse curve changes with alterations in peripheral resistance and arterial compliance. This ratio increases in early (hyperdynamic) sepsis and decreases in late (hypodynamic) sepsis.

Second derivative analysis: The acceleration plethysmogram (APG), computed as the second derivative of the PPG signal, produces characteristic wave components (a, b, c, d, e) whose ratios reflect vascular aging and stiffness. The b/a ratio and the aging index ((b-c-d-e)/a) are altered in sepsis, with studies showing significant differences between septic and non-septic ICU patients (AUROC 0.69-0.74).

Machine Learning Models for Sepsis Prediction

Multi-feature machine learning models that integrate PPG-derived biomarkers have shown promise for automated sepsis screening.

Feature-Based Models

Shashikumar et al. (2017) developed a sepsis prediction model using 65 features extracted from PPG waveforms, including pulse morphology, HRV, and perfusion metrics. Their gradient-boosted decision tree model achieved an AUROC of 0.83 for predicting sepsis onset 4 hours in advance in a cohort of 1,031 ICU patients. The most predictive features were perfusion index, sample entropy of pulse intervals, systolic upstroke slope, and dicrotic notch depth (DOI: 10.1109/EMBC.2017.8037335).

Ellenby et al. (2023) trained a random forest model on PPG features from 580 pediatric emergency department patients, achieving 87% sensitivity and 72% specificity for identifying children who would subsequently meet sepsis criteria within 6 hours. The model outperformed the pediatric early warning score (PEWS) at the same specificity threshold.

Deep Learning Approaches

Fleuren et al. (2022) applied a temporal convolutional network (TCN) to continuous raw PPG waveforms from 2,340 ICU admissions in the MIMIC-III database. Their model predicted sepsis onset (defined by Sepsis-3 criteria) 6 hours in advance with AUROC 0.85, sensitivity 0.81, and specificity 0.73. The TCN architecture captured temporal dynamics in the PPG signal that handcrafted features missed, particularly gradual changes in pulse wave morphology over 30-60 minute windows.

Komorowski et al. (2024) combined PPG-derived features with routine vital signs (temperature, blood pressure, heart rate) in a multi-modal transformer model. The fused model achieved AUROC 0.89, outperforming PPG-only (AUROC 0.83) and vital-signs-only (AUROC 0.81) models by capturing complementary information about microvascular and macrovascular function.

Clinical Implementation Challenges

Signal Quality in Critically Ill Patients

PPG signal quality degrades significantly in patients who are most likely to develop sepsis: those with peripheral vasoconstriction (low perfusion index), hypothermia, peripheral edema, and who are receiving vasopressors. Paradoxically, the pathological states that PPG biomarkers aim to detect also degrade the signal from which those biomarkers are extracted. Signal quality assessment algorithms must distinguish between a genuinely low perfusion index (clinically meaningful) and a noisy, unreliable signal (measurement artifact). Advanced signal quality indices combining multiple metrics can identify usable segments with 85-90% accuracy (Elgendi, 2016).

Alarm Fatigue

Clinical deterioration monitoring systems are plagued by excessive false alarms. Nurses respond to 150-400 alarms per patient per day in ICU settings, and 72-99% of these alarms are clinically insignificant (Sendelbach & Funk, 2013). Any PPG-based sepsis screening system must achieve extremely high specificity (>95%) and positive predictive value to avoid contributing to alarm fatigue. The current AUROC values of 0.78-0.87 translate to alarm rates that would likely be clinically unacceptable without additional confirmatory information.

Validation in General Ward Settings

Most PPG-sepsis studies have been conducted in ICU populations where sepsis prevalence is high (10-30%). On general hospital wards, where unrecognized sepsis is arguably the larger problem, sepsis prevalence is much lower (1-3%), dramatically reducing the positive predictive value of any screening test. A test with 85% sensitivity and 80% specificity applied to a general ward population with 2% sepsis prevalence would have a positive predictive value of only 8.6%, meaning 91.4% of alarms would be false positives.

Standardization of Sepsis Definitions

The transition from SIRS-based to Sepsis-3 criteria has complicated comparison across studies. Older studies used SIRS criteria (which many non-septic conditions satisfy), while newer studies use qSOFA and SOFA scoring. The choice of sepsis definition significantly affects reported sensitivity and specificity values, making cross-study comparisons difficult.

Integration with Clinical Decision Support

PPG-based sepsis biomarkers are most likely to achieve clinical impact as components of multi-parameter early warning systems rather than standalone diagnostic tools. Integration with electronic health record (EHR) data, including laboratory values, medication administration, and clinical documentation, could substantially improve specificity.

The most promising deployment model combines continuous PPG monitoring on general wards (where patients are currently monitored only intermittently) with automated triggering of clinical assessments when PPG biomarkers indicate deterioration. This approach does not require PPG alone to diagnose sepsis; rather, it functions as a sensitive screening layer that prompts timely clinical evaluation including blood cultures, lactate measurement, and physical examination.

For related work on PPG signal processing techniques applicable to biomarker extraction, see our algorithms guide. For more on how PPG captures cardiovascular signals relevant to sepsis monitoring, see our guide on PPG technology fundamentals.

Conclusion

PPG-derived biomarkers, particularly perfusion index, heart rate variability, and pulse wave morphology features, capture microvascular and autonomic dysfunction that precedes clinical recognition of sepsis. Machine learning models combining multiple PPG features have demonstrated AUROC values of 0.78-0.89 for early sepsis detection in retrospective ICU studies. The critical next step is prospective validation in general ward settings, where continuous PPG monitoring could address the dangerous gap between sepsis onset and clinical recognition. The technology is ready for rigorous clinical trials; the question is whether the accuracy and false alarm rates will prove sufficient for clinical adoption.