AI-based Algorithm Development for Improved Diagnostics in the Intensive Care and Acute Stroke Setting Using Electroencephalography

**RATIONALE**

Electroencephalography (EEG) is a noninvasive tool for detecting and monitoring neuronal activity. It plays a crucial role in predicting neurological outcome in postanoxic coma after cardiac arrest (CA). It is also promising for prehospital detection of intracranial hemorrhage (ICrH), including intracerebral hemorrhage (ICH) and subarachnoid hemorrhage (SAH).

Conventional scalp EEG has two practical limitations. Its electrodes must be carefully placed on the scalp by specialized personnel. Its interpretation requires expertise, is time-consuming and is prone to interrater variability.

In critical care and emergency settings, time is critical and specialist availability is often limited. In acute stroke care, rapid detection of ICrH is essential to start early appropriate treatment. In postanoxic coma after CA, early EEG monitoring after ICU admission is recommended to accurately predict neurological outcome.

Recent pilot studies suggest that subhairline EEG is a reliable and easy alternative to scalp EEG for these time-critical settings. Subhairline EEG is a simplified configuration consisting of a single patch applied on the forehead and behind the ears. In addition, artificial intelligence (AI) algorithms for automated interpretation of subhairline EEG could provide rapid and reliable decision support in both settings.

**HYPOTHESIS**

The investigators hypothesize that AI algorithms based on subhairline EEG can accurately predict neurological outcome in patients with postanoxic coma after CA. The investigators also hypothesize that such algorithms can accurately detect ICH in patients with suspected stroke.

**OBJECTIVE**

• Use case 1 (UC1): To develop and validate an AI algorithm based on subhairline EEG for outcome prediction in postanoxic coma at 24 hours after CA.
• Use case 2 (UC2): To develop and validate an AI algorithm based on subhairline EEG and clinical symptoms for detection of ICH in patients with (suspected) stroke.

**STUDY DESIGN**

This is an investigator-initiated, monocenter, observational study at Amsterdam UMC. It consists of three steps.

Step 1: Algorithm development. Several EEG-based AI algorithms will be developed and pre-trained using existing scalp EEG data and clinical data (UC1: >900 patients, UC2: >1000 patients).

Step 2: Prospective data collection. Subhairline EEG data and clinical data will be collected with the TrianecT subhairline EEG recording device. Sample sizes are based on 10-20 events per predictor.

• UC1: 180-360 patients
• UC2.1 (suspected stroke): 300-600 patients
• UC2.2 (suspected SAH): 100 patients

Step 3: Algorithm fine-tuning, validation and feasibility. The pre-trained algorithms will be fine-tuned and validated on the collected subhairline EEG data using transfer learning. The algorithms will then be integrated into the TrianecT software (SMART-EEG). Practical and clinical feasibility will be evaluated in up to 10 patients per use case. This evaluation is not intended for CE certification.

**STUDY POPULATION**

All participants are 18 years or older.

• UC1: Patients in postanoxic coma after CA admitted to the ICU.
• UC2.1: Patients with suspected stroke, as assessed by ambulance personnel, or with confirmed stroke.
• UC2.2: Patients with suspected SAH, as judged by the attending paramedic, or with confirmed SAH.

**INTERVENTION**

All participants undergo a subhairline EEG recording with a single patch applied on the forehead and behind the ears.

• UC1: continuous subhairline EEG recording, in addition to the standard scalp EEG used in routine care. The recording starts as soon as possible after ICU admission and continues up to 72 hours after CA.
• UC2: a single subhairline EEG recording of approximately 2 minutes, performed as soon as possible after presentation at the hospital.

**MAIN END POINTS**

Primary end points:

• UC1: the prognostic performance of subhairline EEG-based AI algorithms for prediction of neurological outcome at 24 hours after CA in postanoxic coma patients in the ICU. This is expressed as the area under the ROC curve (AUC), sensitivity, specificity and predictive values, with 95% confidence intervals. The optimal cut-off for poor outcome is defined as the highest sensitivity at a false positive rate of 0%.
• UC2: the diagnostic accuracy of subhairline EEG-based AI algorithms for detection of ICH in (suspected) stroke patients in the emergency department. This is expressed as the AUC, and as sensitivity, specificity, PPV and NPV at the optimal cut-off, each with 95% confidence intervals.

Secondary end points:

• Prognostic performance of AI algorithms based on subhairline EEG for outcome prediction at each hour between 6 and 24 hours after CA.
• Diagnostic accuracy of AI algorithms based on subhairline EEG and clinical symptoms for detection of SAH.
• Feasibility of SMART-EEG use in the ICU setting.
• Feasibility of SMART-EEG use in the acute stroke setting, including user-friendliness as rated by users.