PPG-Based Delirium Detection in the ICU: Autonomic Signatures and Early Warning

Machine Learning Approaches

Binary Delirium Classification

Classifiers such as random forests can be trained on HRV features from ICU PPG recordings to detect concurrent CAM-ICU-positive delirium, but no validated system with published accuracy is cited here. Performance is expected to vary by subtype. Hyperactive delirium (agitation, combativeness) has more pronounced sympathetic activation. Hypoactive delirium (somnolence, reduced responsiveness) overlaps with normal sleep and sedation patterns.

Hypoactive delirium is the clinically most challenging subtype — it is underdiagnosed by clinical assessment and also the most difficult to detect physiologically.

Predictive Models (Pre-Delirium Detection)

Early prediction of delirium onset — identifying patients who will become delirious in the next 4-8 hours before overt cognitive changes — is more clinically valuable than concurrent detection because it enables prophylactic intervention.

Whether reliable pre-delirium HRV signatures exist is still an open research question.

Multimodal Integration

Combining PPG HRV features with electronic health record data (age, APACHE score, sedation exposure, medications) is expected to improve prediction over HRV alone. EHR risk features (benzodiazepine use, sleep deprivation, pain scores) provide context that pure physiological monitoring lacks.

Practical Implementation Challenges

ICU Signal Quality

ICU patients present extraordinary PPG signal quality challenges:

• Vasopressor infusions: Norepinephrine and vasopressin cause intense vasoconstriction that reduces PPG amplitude. Perfusion index may fall very low, making wrist or finger PPG unreliable. Forehead or ear-canal sensors maintain better signal quality during vasopressor therapy.

• Hemodialysis and continuous renal replacement therapy (CRRT): Blood line connection and disconnection produce large transient artifacts. Dialysis access lines on the same limb as PPG sensors create venous pressure changes affecting DC component.

• Mechanical ventilation: Positive-pressure ventilation produces rhythmic intrathoracic pressure changes that modulate PPG amplitude. These interactions are diagnostically informative (for respiratory-cardiac coupling) but complicate standard HRV interpretation.

Sedation Confounding

Many ICU patients are sedated with propofol, dexmedetomidine, or benzodiazepines. These agents have profound effects on autonomic function and HRV. Dexmedetomidine, which acts via α2-adrenergic receptors, produces a distinct HRV pattern (reduced LF/HF, slowed heart rate) that must be distinguished from delirium-related autonomic changes.

Sedation-aware models incorporate Richmond Agitation-Sedation Scale (RASS) scores and sedation medication exposure to reduce misclassification.

Toward Continuous Delirium Monitoring

The clinical implementation pathway for PPG-based delirium monitoring involves:

1. Background physiological baseline: 24-hour PPG recording during ICU admission provides individual baseline HRV parameters
2. Continuous monitoring: Hour-by-hour delirium probability scores updated in real time in the EMR
3. Alert thresholds: Probability exceeding 0.75 triggers nursing assessment with CAM-ICU
4. Intervention logging: Treatment responses (reducing sedation, adding haloperidol, mobilizing patient) logged with timestamps for algorithm learning

This workflow transforms delirium monitoring from periodic snapshots to continuous surveillance, matching the actual clinical course of a fluctuating syndrome.

FAQ

What causes autonomic dysfunction in delirium? The autonomic changes in delirium result from multiple converging mechanisms: systemic inflammation activates the sympathetic nervous system via cytokine signaling; cholinergic neurotransmission is impaired (reducing vagal tone); circadian rhythm disruption from the ICU environment (constant light, noise, night-time interventions) abolishes normal day-night autonomic cycling; and underlying critical illness (sepsis, cardiac failure, respiratory failure) independently drives autonomic dysregulation.

How does ICU delirium differ from normal ICU sedation in PPG terms? Both reduce HRV overall, but delirium specifically disrupts circadian HRV patterns and produces unstable beat-to-beat variability (reduced complexity measures), whereas appropriate sedation produces more predictable, structured HRV reduction with preserved phase responses. Dexmedetomidine sedation produces a particularly distinctive HRV pattern that must be specifically accounted for in delirium detection algorithms.

Can PPG delirium detection distinguish hyperactive from hypoactive delirium? Hyperactive delirium (agitation) produces greater sympathetic activation (higher LF power, higher LF/HF ratio, more pronounced tachycardia) and is detected with higher accuracy than hypoactive delirium. Hypoactive delirium, which is harder to identify clinically and physiologically, remains the more challenging detection target.

Is continuous PPG monitoring feasible for all ICU patients? Feasibility depends on hemodynamic status. Patients receiving vasopressors have severely reduced peripheral perfusion that degrades wrist/finger PPG signal quality. Forehead, earlobe, and esophageal PPG sensors maintain acceptable signal quality in more compromised patients.

What is the clinical benefit of detecting delirium earlier? Earlier delirium detection enables earlier non-pharmacological interventions: promoting sleep-wake cycle, reducing unnecessary sedation, mobilization, family engagement, noise reduction, and reorientation. These ABCDEF bundle interventions are more effective when applied proactively. Antipsychotics are not a substitute: the MIND-USA trial found that haloperidol and ziprasidone did not increase days alive without delirium or coma compared with placebo.

How accurate are current PPG-based delirium detection models? PPG-based delirium detection is still a research area, and no validated system with published accuracy is cited here. By comparison, the nurse-administered CAM-ICU had sensitivities of 93-100% and specificities of 98-100% in its validation study, but it only captures the moment of assessment.

References

1. Ely, E.W., Inouye, S.K., Bernard, G.R., Gordon, S., Francis, J., May, L., et al. (2001). Delirium in mechanically ventilated patients: validity and reliability of the confusion assessment method for the intensive care unit (CAM-ICU). JAMA, 286(21), 2703-2710. doi:10.1001/jama.286.21.2703

2. Girard, T.D., Exline, M.C., Carson, S.S., Hough, C.L., Rock, P., Gong, M.N., et al. (2018). Haloperidol and ziprasidone for treatment of delirium in critical illness. New England Journal of Medicine, 379(26), 2506-2516. doi:10.1056/NEJMoa1808217