Systematic Review on Noncomputed Tomography‐Based Technologies for Prehospital Stroke Diagnosis

Abstract

Effects of stroke therapies area highly time dependent but onset‐to‐treatment times for recanalizing treatment are mostly beyond optimal time windows. Mobile stroke units shorten time to treatment and improve functional outcome. However, costs and logistical demands preclude their widespread use. There is a pressing need for alternative prehospital diagnostic tools that enable early, on‐site treatment and facilitate targeted transport in patients needing interventional therapies. This systematic review aimed at examining the current landscape of noncomputed tomography‐based technologies for prehospital stroke diagnosis. We conducted a literature search of original research papers on predetermined technologies and through systematic scouting on PubMed and included papers in English found through MEDLINE and Embase, published until October 2025. Relevant papers not identified by the systematic search but known to the authors were included and indicated as such. The final review included 105 publications evaluating innovations in microwave, transcranial eddy current damping, volumetric impedance phase shift spectroscopy, ultrasound, magnetic resonance imaging, electroencephalography, near‐infrared spectroscopy, blood biomarkers, telemedicine, and artificial intelligence. Each technology's potentials, limitations, and clinical feasibility for enhancing prehospital diagnosis and therapeutic and monitoring utility were assessed. No single technology provides capabilities similar to noncomputed tomography in terms of detection and differentiation of ischemic and hemorrhagic stroke. A combination of ultrasound and near‐infrared spectroscopy may complement each other in hemorrhage detection. Electroencephalography shows most promise in identifying large vessel occlusion. Telemedicine can be used for reliable prehospital stroke assessment. Future research should target both the improvement of individual technologies as well as combining different modalities.

Nonstandard Abbreviations and Acronyms

ECD

eddy current damping

GFAP

glial fibrillary acidic protein

ICH

intracerebral hemorrhage

LVO

large vessel occlusion

MSU

mobile stroke unit

MWI

microwave imaging

NIRS

near‐infrared spectroscopy

Reperfusion therapies like intravenous thrombolysis and mechanical thrombectomy significantly improve functional outcomes after acute ischemic stroke, but their effectiveness relies on swift treatment. ¹ , ² Studies have shown a dramatic improvement in outcomes when treatment occurs within 60 minutes of onset, the “Golden Hour,” ³ yet median onset to treatment times for in‐hospital thrombolysis typically exceed 100 minutes. Mobile stroke units (MSUs) with onboard computed tomography (CT) scanners were developed to shift stroke workup to the prehospital setting, and recent studies confirm that MSUs not only shorten time to treatment but also improve outcomes. ⁴ This has resulted in a class 1 recommendation in the latest guideline of the American Heart/Stroke Association. ⁵ However, MSUs are costly, pose logistical challenges, and rely on precise dispatching, limiting their scalability particularly outside of metropolitan areas. Hence, stroke diagnostic technologies designed for usage in normal ambulances could leverage the prehospital treatment concept for a much higher proportion of patients with stroke in the field.

Although mechanical thrombectomy remains an in‐hospital procedure at specialized centers, delays are common due to secondary transfers from facilities without endovascular capabilities. Ambulance‐based stroke workup with vessel imaging could identify patients with large‐vessel occlusions (LVOs) and route them directly to mechanical thrombectomy‐capable centers. Prehospital diagnosis of stroke subtype and vessel occlusion could also allow hospitals to be prenotified, speeding up treatment, enabling enhanced thrombolytic therapy (eg, sonothrombolysis), and facilitating hemostatic treatment for patients with intracerebral hemorrhage (ICH). This systematic review outlines the current status of non‐CT‐based stroke diagnostics for prehospital use complemented by critical appraisals from researchers involved in these technological developments.

METHODS

Search Strategy

In order to determine whether important emerging technologies had been missed in our initial scouting, we conducted a systematic search using the terms (“Stroke”[Mesh] OR stroke OR “Ischemic Stroke”[Mesh] OR “Ischemic Stroke” OR “Hemorrhagic Stroke”[Mesh] OR “Hemorrhagic Stroke” OR “Cerebral Hemorrhage”[Mesh] OR “Cerebral Hemorrhage”) AND (“pre‐hospital” OR prehospital OR “Emergency Medical Services”[Mesh] OR “emergencies”[Mesh] OR “ambulances”[Mesh] OR ambulance) AND (“Stroke/diagnosis”[Mesh] OR “Stroke/diagnostic imaging”[Mesh]) in PubMed. Headlines were screened for technologies and approaches. Next, we conducted a systematic review by entering the comprehensive search terms for every technology in PubMed and Embase (Supplemental Material 1). Databases were last consulted on October 30, 2025. Literature not identified through the systematic search but already known by the authors was also considered. For each technology, 2 authors independently reviewed titles and abstracts for eligibility. Discrepancies were adjudicated by an independent referee. Finally, full texts were reviewed by 2 independent reviewers.

Selection Criteria

After removal of duplicates, our initial search identified 7609 database entries. Studies were included in the final review if they were original articles, written in English language, describing technologies and processes for prehospital stroke diagnosis. The final analysis included 136 studies (Figure1). All technologies were reviewed according to predefined criteria (Table1).

Figure 1. Preferred Reporting Items for Systematic Reviews and Meta‐Analyses flow diagram for article identification.

Searches were conducted in Medline via PubMed and Embase. Additional sources were articles already known to the authors.

Table 1.

Technology Rating Matrix

	Microwave imaging	ECD sensors	VIPS	EEG	MRI	NIRS	Biomarkers	Ultrasound	Telemedicine	AI
Needed space and weight	++	++	++	++	−	++	++	++	+	++
Ability to differentiate stroke and stroke mimics	+	−	+	+	++	−	+	−	++	++
Arterial vessel imaging	−	−	−	−	−	−	−	++	−	−
Perfusion imaging	−	−	−	−	+	+	−	+	+*	+*
(Indirect) evaluation of large vessel occlusion	+	+	+	+	+	+	+	++	++	+
Detection of deep intracerebral hematoma	+	+	+	−	+	−	+	++	−	+
Detection of superficial hematoma	+	+	+	−	+	++	+	−	−	+
Spatial resolution	+	+	−	+	++	+	−	++	+	+
Blind spots in supratentorial cavity	−	−	−	+	−	+	−	+	−	n.a.
Blind spots in posterior fossa	+	+	−	++	+	++	−	++	−	−
Susceptibility to motion artifacts	−	−	−	++	+	−	−	+	−	−
Potential therapeutic utility	−	−	−	−	−	−	−	+	+	+
Potential monitoring utility	+	−	−	+	−	+	+	+	++	+

− indicates Week; +, Middle; ++, Strong; AI, artificial intelligence; ECD, eddy current damping; EEG, electroencephalography; MRI, magnetic resonance imaging; n.a., not applicable; NIRS, near‐infrared spectroscopy; and VIPS, volumetric impedance phase shift spectroscopy.

^*By indirect assessment using clinical severity.

RESULTS

Microwave Imaging

Microwave Imaging (MWI), also termed electromagnetic brain scanning, is a portable and low‐cost diagnostic tool that uses tissue dielectric variations to differentiate between healthy and abnormal tissues. ⁶ Initially used for breast cancer detection, MWI has shown potential for stroke diagnosis. ⁷ , ⁸ , ⁹ , ¹⁰ MWI works by transmitting microwave signals through the brain using distributed antennas around the head for enhanced sensitivity. ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ Each antenna serves as a transmitter one at a time, and the remaining antennas act as receivers. ⁸ The scattered wave pattern resulting from variations in dielectric contrast are analyzed to differentiate between healthy tissue and pathological substrates. ⁷ , ⁸ , ⁹ , ¹¹

Two main approaches, microwave tomography and ultra‐wideband radar imaging, are employed. ⁶ Microwave tomography reconstructs tissue morphology and electromagnetic characteristics but requires solving complex inverse problems and a training phase. ⁶ Ultra‐wideband radar imaging locates scattering points, generating energy maps. ⁶ Computational advances in the past 5 to 10 years have made it clinically viable. ⁷ , ⁸

Four MWI devices studies had published scientific evaluations ¹² : EMTensor (EMTensor, Inc., Newport Beach, CA, and EMTensor, GmbH, Vienna, Austria), EMvision (EMvision Medical Devices, Ltd., Brisbane, Australia), StrokefinderTM MD100 (Medfield Diagnostics AB, Göteborg, Sweden), ¹² , ¹³ and TES HT100 (B&B Srl, Naples, Italy). ¹⁴

Ljungqvist et al. demonstrated MWI's effectiveness in detecting subdural hemorrhage, achieving an area under the curve (AUC) of 0.94 with 100% sensitivity and 75% specificity. ⁹ Persson et al. studied 9 patients with ICH and 11 patients with acute ischemic stroke and showed MWI could distinguish between the 2 with an AUC of 0.88. In a subsequent phase of the study, it correctly differentiated between ICH and IS (AUC, 0.85) and ICH and healthy volunteers (AUC, 0.87). ⁸ Abbosh et al. studied 50 patients with ischemic or hemorrhagic stroke and indicated promising results in detecting, localizing, and classifying strokes. ¹⁵

Montemarani et al. examined 23 patients with mild traumatic brain injury after hospital arrival and compared the results of a 5‐minute portable TES HT 100 device analysis with findings in cranial CT. After exclusion of 3 patients with storage error, they found complete concordance with regard to any brain lesion (14 patients with and 6 patients without brain lesion). No information was provided for the type or location of the brain lesion. ¹⁴

Transcranial Eddy Current Damping Sensors

Transcranial eddy current damping (ECD) sensing is a fairly recent, experimental approach to locating, distinguishing hemorrhagic and ischemic strokes. ¹⁶

The method uses a solenoid magnet to induce eddy currents in the radio‐frequency range within a tissue, whose magnitude depends on electric conductivity, determined by the tissue's composition. The resulting electromagnetic field reduces the inductance of the coil, dampening resonance in an oscillator circuit that is measured in real time.

The ECD device consists of a handheld probe containing ≥1 coils of different sizes and inductances, an integrated circuit that drives the coil and detect inductance changes, and a microcontroller. Data are collected by moving the probe over the head, transmitting the measurements to a computer for 2‐dimensional or 3‐dimensional mapping. Larger coils provide greater sensing depth but lower spatial resolution. ECD identifies lesions by sensing irregularities in tissue conductivity: hemorrhages increase conductivity due to ionic concentration (not restricted by cell membrane) in the hematoma, and ischemic lesions show reduced conductivity.

A study using an anthropomorphic head phantom achieved centimeter resolution in simulated hemorrhages of 30 to 50 mL, requiring 2.43 minutes for a complete scan. ¹⁶

A clinical study with eight subjects (3 hemorrhagic strokes, 2 ischemic, 3 healthy) demonstrated accurate lesion localization and differentiation among the stroke types within an average scan time of 15 minutes. ¹⁷ The investigators were blind with respect to the location and type of lesion, and all lesions were correctly localized and identified. The method showed greater sensitivity to hemorrhages than ischemic lesions. Further phantom tests estimated a sensing depth of up to 4.97 cm, potentially covering most of the human cranium. However, ECD's ability to differentiate intracerebral hematomas from other fluid‐filled structures like ventricles, arachnoid cysts, or lesions deeper than 5 cm remains uncertain. In a recent phantom study, the addition of machine learning (ML) further improved classification metrics. ¹⁸

Volumetric Impedance Phase Shift Spectroscopy

Volumetric impedance shift spectroscopy leverages bioimpedance when the distribution of fluids between compartments shifts in tissue affected by stroke. An array of low energy radio waves is sent from the back of a portable, helmet‐like device to the front, thereby being modified when passing through brain tissue. ¹⁹ In a study with 252 patients and healthy controls, a volumetric impedance phase shift spectroscopy device was able to differentiate subjects with severe stroke from the other participants with an AUC of 0.93 (95% CI, 0.89–0.96). ¹⁹

Electroencephalography and Somatosensory Evoked Potential

Electroencephalography (EEG) and somatosensory evoked potential have been addressed for prehospital stroke detection in various studies. EEG parameters tested included theta/alpha ratio, relative theta power, weighted phase lag index, ²⁰ , ²¹ delta/alpha ratio, (delta + theta)/(alpha + beta) ratio, pairwise‐derived brain symmetry index, ²² fast frequency amplitudes, changes in relative delta power, or indices derived from more complex algorithms. ²³ Regarding somatosensory evoked potentials, 1 study used complex automated algorithms. ²³

Most studies focused on LVO detection, with only 1 conducted prehospital. ²¹ Although some studies explicitly restricted the examined area to the anterior circulation, ²⁰ , ²¹ , ²³ even in the remaining studies, no LVO was detected in the posterior circulation. ²⁴

Accuracy varied: Sergot et al. reported 80% sensitivity and specificity using a portable EEG/somatosensory evoked potential device. ²³ The ELECTRA (EEG Controlled Triage in the Ambulance for Acute Ischemic Stroke) study showed in patients with analyzable EEG measurements (212/311 [68%]) an AUC of 0.80 for the theta/alpha ratio, with 50% sensitivity and 83% specificity. The Brain Symmetry Index in the delta band had an AUC of 0.91, with 80% sensitivity and 93% specificity. ²¹ In an emergency department (ED) study by the same group, combining theta power and weighted phase lag index yielded 100% sensitivity and 84% specificity for LVO detection in 65 patients with suspected stroke (ischemic or hemorrhagic) or confirmed LVO. ²⁰

Other studies focused on stroke detection per se, and the definition of stroke varied. Wilkinson et al. tested a consumer wireless EEG system in patients with subacute ischemic stroke against healthy controls and stroke mimics. ²² Delta/alpha ratio and (delta + theta)/(alpha + beta) ratio increased with stroke severity with significant differences between large strokes and both small strokes and controls. A model differentiating between minor strokes/controls and moderate/severe strokes had a 76% accuracy with a sensitivity of 63% and a specificity of 86%.

Michelson et al. compared EEG to CT in and patients with acute ischemic stroke and ICH and stroke mimics (migraine and syncope). Sensitivity of stroke detection was 91.7%, specificity 50.4%. Eighty percent of patients with ischemic stroke with normal CT findings were correctly identified as stroke by EEG at the time of CT scan, whereas no differentiation between ischemic and hemorrhagic stroke was attempted. ²⁵

Portable EEG devices, especially dry electrode cap systems, reduced setup times to around 5 minutes, with 1‐hour training sufficient for examiner‐independent results. ²⁰ , ²¹ , ²² , ²³ One study conducted in the prehospital setting in patients with neurological symptoms reported easy application and EEG recordings usable for clinical interpretation but not specifically for stroke diagnosis. ²⁶

The primary limitation of all studies using EEG is the high incidence of poor signal quality recordings, which consequently results in a high dropout rate. Numbers ranged between 10% and 34%. ²⁰ , ²³ , ²⁷ This was particularly true for women, patients with long hair, and those with LVOs, who comprised the target population for the studies. ²⁰ , ²¹ None of the studies differentiated LVO from hemorrhagic stroke, precluding early recombinant tissue plasminogen activator administration in prehospital settings. Deep brain regions remain challenging for accurate detection.

Portable Magnetic Resonance Imaging

Magnetic resonance imaging (MRI) is a noninvasive technology that uses a magnetic field and radiofrequency pulses to create detailed images of the brain including the posterior fossa, which is challenging for many other technologies. MRI is highly sensitive for detecting acute ischemic stroke, with diffusion‐weighted imaging showing changes within minutes of arterial occlusion. However, traditional MRI machines require large magnetic fields, advanced infrastructure, and cooling with cryogens. Recent developments have led to portable MRI scanners with ultra‐low field strength (<0.1 T), which are compact (140×90 cm), lightweight (650 kg), and operate on standard electrical outlets without the need for cryogens. ²⁸ , ²⁹ , ³⁰ These devices are considerably cheaper, making them potentially more accessible even for low‐income communities in the future. ³¹ , ³² Portable MRI is currently most promising in critical care settings, enabling bedside scans for mechanically ventilated patients. However, caution is warranted in patients with pacemakers, cochlear implants, or deep brain stimulators. ³³

For prehospital stroke care, portable MRI faces 2 main challenges; long image acquisition times and limited image quality. Current ultra‐low field scanners do not support perfusion imaging or time‐of‐flight angiography, essential for identifying patients needing endovascular therapy. Other low‐field (eg, 0.5 T) MR scanners with time‐of‐flight angiography capability are currently not portable (≈2000 kg). ³² , ³⁴

Few studies have examined the use of portable MRI in stroke care. One prospective single‐center study in 2019 to 2020 included 50 critically ill patients with various conditions including ischemic stroke (n=9), hemorrhagic stroke (n=12), brain tumor, and brain injury. Portable MRI findings largely matched conventional imaging (CT or MRI), except for 1 missed subarachnoid hemorrhage, which could have had serious implications for stroke treatment. ³³ In a subsequent study involving 144 individuals (56 ICH, 48 acute ischemic stroke), portable MRI showed 80.4% sensitivity for ICH, missing 11 cases. The lack of T2* and susceptibility‐weighted imaging likely contributed to this limitation. Specificity was high at 96.6%, with 3 patients misclassified as having ICH. ²⁹ More recently, portable MRI was tested in 50 patients with subacute ischemic stroke. The diffusion‐weighted imaging sequence, critical for hyperacute stroke detection, had modest quality, showing 86% sensitivity. Five patients with small infarcts (4–10 mm) were missed. T2‐W and fluid‐attenuated inversion recovery sequences had higher sensitivity at 98% and 100%, respectively. ³⁵ This portable MRI system (Swoop, Hyperfine, Inc., Guilford, CT) was the first to gain Food and Drug Administration approval in August 2020. It was successfully integrated into a modified cargo van, showing similar imaging quality to static systems. ³¹

Image acquisition times of 7 to 10 minutes are too slow for the hyperacute stroke setting. Sequences with long acquisition time increase the risk of motion artifacts, especially in agitated patients with stroke. Several teams are working to improve acquisition times and image quality using deep learning and undersampling. ³⁶ , ³⁷ Rigorous clinical studies are needed to assess whether faster acquisition approaches do not increase false negative and false positive findings.

Near‐Infrared Spectroscopy and Other Optical Technologies

Near‐infrared spectroscopy (NIRS) is a noninvasive, easy‐to‐apply technique that has been used for years to observe cerebral oxygenation by measuring the absorption of 690 to 830 nm light by erythrocytes approximately in a depth of up to 2 to 3 cm. NIRS has the potential capability to detect superficial cranial bleeds, such as subdural and superficial intracerebral hematomas or subarachnoid hemorrhages over the brain convexity. Depending on the device used, blood and blood oxygenation may be computed in 1 or in several locations simultaneously. The devices can be battery operated and are portable to a varying degree, making them potentially attractive for prehospital diagnostics. The few human studies have focused largely on traumatic intracranial hemorrhage and not on stroke. ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴² , ⁴³

In 1 study by Gopinath et al., hemispheric differences were observed in 167 traumatic brain injury cases, suggesting potential to identify hemorrhage, although controls were absent, leaving sensitivity and specificity unverified. ³⁸ In a multicenter study of 365 trauma patients, NIRS showed 88% sensitivity and 90.7% specificity for intracranial hematomas >3.5 mL and within 2.5 cm from the skull's surface. However, for all hematoma locations, sensitivity fell to 68.7% with the same specificity. ³⁹

Moreau et al. studied regional NIRS changes in 5 patients with stroke versus cadavers and healthy controls, observing differences that suggested NIRS may monitor stroke evolution, though data quality issues arose. ⁴¹

Muehlschlegel et al. evaluated interhemispheric asymmetry in 9 patients after ischemic stroke compared with controls, proposing NIRS might detect hemispheric microvascular hemodynamic asymmetry relevant to prehospital stroke assessment. ⁴²

Similar to NIRS, 1 study used an optical device that illuminates tissue with high‐intensity laser light and measures speckle contrast to monitor cerebral blood flow. ⁴⁴ In a cohort of 135 patients, the device was able to detect LVO with a sensitivity of 78.9% (95% CI, 65.5%–88.0%) and a specificity of 84.3% (95% CI, 74.7%–90.7%). ⁴⁴

In summary, modern NIRS devices are portable and user friendly, yet high‐quality evidence for prehospital stroke diagnosis remains limited. Most studies did not address ischemic stroke, and applicability to early stroke is unclear. The studies are based on small case series, few provide matched control groups, and due to a significant risk of bias and imprecision, the level of current evidence is weak. Specificity and sensitivity in stroke identification have not been investigated and many other questions like device resilience to motion and other artifacts are yet to be addressed.

Blood‐Based Biomarkers

Innovative approaches like integrating blood‐based biomarkers in ambulances offer potential in prehospital stroke diagnosis. Despite many years of research, no biomarkers have become standard in clinical practice, but advancements in proteomics, metabolomics, transcriptomics, and genomics have led to new candidate biomarkers. ⁴⁵

A proof‐of‐concept study from Helsinki showed that obtaining blood samples in ambulances could save around 30 minutes compared with in‐hospital sampling, with point‐of‐care measurements enabling earlier treatment decisions. ⁴⁶

An MSU study found several elevated neuroinflammatory factors in patients with stroke only minutes after onset, supporting further exploration of ultra‐early immune response in stroke. Although promising, the sample size was small, and further studies are needed. ⁴⁷

Attempts to differentiate stroke from mimics using biomarkers have largely been unsuccessful. In fact, a 21‐biomarker panel selected among previous studies and from the literature showed that the studied biomarkers were not sufficient for an accurate differential diagnosis of stroke. ⁴⁸ Copeptin, a general outcome predictor, did not show discriminatory value between stroke and stroke mimics in the prehospital setting. ⁴⁹ Recently, small noncoding RNAs have been identified as potential candidates; however, RNA sequencing is not yet feasible in the prehospital setting. ⁵⁰ Symmetric dimethylarginine shows certain promise in severe stroke patients, but only in late sampling windows. ⁵¹

Among the most validated markers for prehospital use, GFAP (glial fibrillary acidic protein) shows potential in identifying ICH. Studies in hospital settings demonstrated GFAP's high sensitivity and specificity for diagnosing ICH. In a prehospital setting, Rozanski et al. collected plasma samples from ICH and ischemic stroke patients and measured GFAP using a fully automated immunoanalyzer platform, reporting a sensitivity of 36% and a specificity 100% for identifying ICH using a GFAP cutoff of 0.29 ng/mL, suggesting its potential for treatment initiation in patients with ICH. ⁵²

Another study using ELISA reported that GFAP accurately (AUC, 0.86) differentiated ICH from ischemic stroke and mimics, regardless of age, severity, or time elapsed since onset, with GFAP being correlated to ICH volume. ⁵³ Another prehospital study showed that serial prehospital GFAP measurements improved ultra‐early differentiation of stroke subtypes, while there was no additional value of the biomarker TAU in ICH identification. ⁵⁴

Studies using a point‐of‐care device for GFAP measurements demonstrated similar predictive values, streamlining prehospital measurements. ⁵⁵ , ⁵⁶

In the Treat‐NASPP (Treat‐Norwegian Acute Stroke Prehospital Project), serum GFAP levels were measured in 299 patients with suspected stroke and onset ≤4.5 hours. ⁵⁷ A decision tree combining GFAP serum levels with a prehospital stroke score was developed to differentiate between ICH, ischemic LVO/non‐LVO, and stroke mimics. This approach first identified patients with a low stroke probability, then used GFAP to detect ICH with 25% sensitivity and 100% specificity.

In ED studies, the diagnostic accuracy improved by adding ischemia‐related biomarkers like RBP‐4 (retinol binding protein 4) or NT‐proBNP (N‐terminal pro‐B‐type natriuretic peptide) to GFAP. If these findings are replicated in ambulances and rapid GFAP tests are developed, biomarker‐based stroke workups could enable targeted triage of patients with ICH (eg, directing them to hospitals with neurosurgery) and potentially allow for prehospital ICH treatment studies.

In a retrospective study, endogenous tissue‐type plasminogen activator levels were lower in patients with hemorrhagic stroke, and their inclusion improved classification accuracy when combined with prehospital stroke scales and vital parameters. ⁵⁸

The ongoing BIOSHIP (Potential Use of a POCT [Point‐of‐Care Test] for the Direct Transfer of Stroke Patients From the Ambulance to the Thrombectomy Center) study (https://clinicaltrials.gov/study/NCT05833945) is evaluating a rapid point‐of‐care test for FABP (fatty acid binding protein) and NT‐proBNP in ambulances to detect LVO patients. ⁵⁹ The addition of D‐dimer to FABP and NT‐proBNP might become a promising LVO biomarker panel. ⁶⁰

Although blood‐based biomarkers in ambulances show promise for stroke diagnosis, challenges remain, including assay standardization, device availability, and emergency medical services (EMS) personnel training.

Transcranial Ultrasound

Transcranial ultrasound is a promising tool for stroke detection, primarily focusing on LVO and intracranial hemorrhage detection. ⁶¹ , ⁶² , ⁶³ , ⁶⁴ , ⁶⁵ , ⁶⁶ , ⁶⁷ , ⁶⁸

Studies show that experienced stroke neurologists using transcranial color‐coded duplex sonography can detect LVO, especially in the middle cerebral artery, with sensitivity up to 91% and specificity up to 98%. ⁶³ , ⁶⁴ , ⁶⁶ Most studies used microbubble contrast agents, improving sensitivity and reducing insufficient bone window rates. Connolly et al. reported only 6% insufficiency using ultrasound contrast agent (SonoVue), ⁶³ whereas typical rates of insufficient bone windows are around 20% in patients with stroke. ⁶⁹ One case report used mobile brain perfusion ultrasound, a technique creating hemispheric perfusion curves following ultrasound contrast injection, to detect middle cerebral artery occlusion in a patient with acute stroke. ⁷⁰

Antipova et al. reported a sensitivity of 63% for hemorrhage detection in a small cohort (10/18 patients). ⁶² A more recent study identified intracerebral hemorrhages in 87.2%. ⁷¹ Localization of the hemorrhages might be more critical than size. This is especially true for cortical and deep infratentorial hemorrhages, albeit Kukulska‐Pawluczuk et al. were able to detect all hemorrhages located in the brainstem and cerebellum. ⁷¹

Ultrasound examinations are quick, with times between 2 and 6 minutes depending on scope (eg, vessel focus, contrast use). ⁶³ , ⁶⁵ , ⁶⁶

With its easy portability, devices can be used at the patient's side, in ambulances, or even helicopters. Perfusion imaging was not addressed in these studies, but it could enhance future stroke diagnostics.

Ultrasound is examiner dependent, which limits its use on ambulances at the scene. Connolly et al. observed significant examination time differences between novice and expert sonographers. ⁶³ Experience also likely affects detection accuracy for hemorrhages and occlusions. This could be addressed by having trained specialists evaluate ultrasound images via telemedicine, with relatively novice ambulance staff handling the scanning. In a study conducted by Antipova et al., novice ultrasound users received a 3‐hour training and were afterwards able to perform simple B‐Mode and transcranial color‐coded duplex sonography examinations on healthy volunteers with satisfactory results. ⁶¹ Mort et al. demonstrated a multichannel communication device that livestreams ultrasound images to hospital‐based experts, though broadband coverage remains a challenge in rural areas. ⁷² Automated or robotic assisted transcranial color‐coded duplex sonography devices may be an alternative to facilitate the use of ultrasound in normal ambulances. The biggest challenge remains the reliable exclusion of intracranial bleeding, which is necessary for the administration of systemic thrombolysis therapy. Smaller, superficial, and infratentorial hemorrhages are particularly challenging for ultrasound‐detection. Experimental procedures, such as the application of ultrasonic shear wave elastography, could possibly improve the negative predictive value in the future. ⁷³ , ⁷⁴

Telemedicine

Telemedicine using video examination enables reliable remote assessment of stroke related neurological deficits. ⁷⁵ , ⁷⁶ , ⁷⁷ , ⁷⁸ Given the time‐dependent effects of intravenous thrombolysis, telemedicine‐based prehospital assessments have been proposed as a means of quality improvement since the early times of telemedicine stroke care. ⁷⁹ Initially, limited bandwidth bore the risk of slow or unstable data transmissions. ⁷⁹ , ⁸⁰ , ⁸¹ With the advent of 4G networks, remote examination before or during EMS transports became more reliable and multichannel connectivity solutions further improved performance. ⁸⁰ , ⁸² , ⁸³

Technical failures during video examination have been reported in most studies, ranging between 60% ⁸¹ when only 3G was available to only 7% in a recent report from New Zealand. ⁸⁴

Telemedicine has also been used for transmitting data from CT imaging ⁸⁵ on an MSU and from other diagnostic devices such as ultrasound. ⁷²

Several studies investigated the feasibility and reliability of simplified neurological assessments ⁸⁶ and examinations without assistance in driving ambulances, ⁸⁷ often using standardized scenarios ⁸⁶ or actors simulating acute stroke. ⁷⁷ , ⁸¹ , ⁸² , ⁸⁷ , ⁸⁸ , ⁸⁹ Usually, comparisons between onsite and remote assessments performed during ambulance transport ⁸¹ , ⁸² , ⁹⁰ showed moderate (weighted kappa 0.4–0.5) ⁸⁹ or moderate to good (weighted Ƙ=0.69) ⁸¹ reliability.

Real‐life reliability has been tested only on MSUs comparing onsite assessment by the MSU neurologist and telestroke examination. ⁹¹ Interrater reliability was high for the National Institutes of Health Stroke Scale sum scores (Ƙ=0.86). Generalizability was limited by the fact that neurological assessments were conducted at the scene and not en route.

Satisfaction rates of prehospital telemedicine users were rated high (>80%) ⁹² and audio/video quality rating improved with higher bandwidth available in 4G and 5G telecommunication standards. ⁷⁷ , ⁸¹ , ⁸² , ⁹¹

Studies report varying impacts of prehospital telestroke on process times. In Germany, early video assessments by nonneurologists in real patients with stroke showed no significant differences in pre‐ and in‐hospital times. ⁸⁰ Duration of neurologists' telemedical assessment ranged between 8 and 19 minutes for the examination at the scene in MSUs ⁸⁵ , ⁹¹ and 18 minutes during transport. ⁸⁹ Hospital prenotification via paramedic‐based stroke assessment and checklist documentation was associated with 4 minutes shorter door‐to‐needle times and higher percentage of door‐to‐needle times ≤20 or 60 minutes in the German Stroke Angel project. ⁹³ Telephone‐based assessment by neurologists or purpose‐trained residents was associated with no difference in onset‐to‐needle times for intravenous thrombolysis but a 69‐minute shorter onset to puncture time for thrombectomy compared with a preimplementation historical cohort in Stockholm. ⁹⁴

The implementation of a joint paramedic and remote physician video examination on ambulances was associated with time saving in hospital‐arrival‐to‐CT time of 5.9 minutes. ⁹⁵ A 17‐minute shorter door‐to‐needle time and 87‐minute shorter onset‐to‐groin time compared with “standard telestroke” was reported after extension of the telemedicine network using telemedicine in EMS assessment in South Carolina. ⁹⁶

Decision accuracy has generally been reported as high in several studies. ⁸² , ⁸⁴ , ⁸⁵ , ⁹¹ , ⁹⁷

Paramedic‐based stroke assessment and telephone‐based consultation by hospital stroke physicians showed 87% accuracy in predicting LVO stroke in the Stockholm Stroke Triage System.

A cluster randomized trial in New Zealand showed better accuracy in identification of reperfusion candidates (intravenous thrombolysis and endovascular thrombectomy) by neurologist‐guided telestroke assessment en route to hospital (100%) than a large‐vessel score performed by paramedics (71%). ⁸⁴

In an adjusted comparison with historical controls, modified Rankin Scale score 0 to 1 in thrombectomy‐eligible patients significantly improved after implementation of the Stockholm Stroke Triage Project. ⁹⁸

Only a small proportion of disagreements regarding stroke and nonstroke diagnosis was reported when comparing onsite and remote (video‐based) assessment of acute patients with suspected stroke on an MSU by neurologists. ⁹¹ However, a German study found that remote, nonneurologist assessments were 91% accurate for stroke diagnosis, there was a 68% misdiagnosis rate for epileptic seizures, frequently misidentified as stroke. ⁹⁹

Artificial Intelligence

The increasing availability of digital neuromonitoring, new sensor, and wearable technologies as well as computer power may offer opportunities to improve prehospital management of acute stroke. Artificial intelligence methods play a crucial role in this process in order to extract the relevant information from these data throughout the stroke care pathway, including diagnosis, consultation, patient transfer, treatment and rehabilitation. ¹⁰⁰ Artificial intelligence methods based on parameters from various sensor systems, including accelerometers, gyroscopes, magnetometers, computer vision systems, audio, pressure sensors, mechanical and optical keyboard press sensors, and microwaves have been explored for stroke detection and management. These technologies aim to detect biometric markers such as speech, motor skills, gait, and sensory changes to identify early signs of stroke. ⁷ , ¹⁰¹ , ¹⁰² , ¹⁰³

Artificial intelligence can also improve prehospital triage by automated speech recognition software. ¹⁰⁴ ML models have demonstrated significantly better prediction performance in identifying patients with stroke in EDs using factors like presenting symptoms, triage level, medical history, vital signs, and presence of atrial fibrillation. ¹⁰⁰ , ¹⁰² , ¹⁰⁵ , ¹⁰⁶ More recently, decision‐support models have also been proposed in the form of smartphone applications to optimize EMS stroke triage, potentially reducing hospital arrival times. ¹⁰⁷

Different modalities and artificial intelligence approaches have been investigated for their potential to detect and classify stroke symptoms. Video‐based approaches showed promise of automatically assessing orofacial stroke impairments. ¹⁰⁸ Other motor symptoms may also be detected using sensors and ML. One validation study assessed motor weakness and graded National Institutes of Health Stroke Scale and Medical Research Council scores of 4 limbs similar to clinical examinations performed by medical staff. ¹⁰³ Specifically, the proposed system composed of a measuring unit with wearable sensors and a grading unit trained with ML was able to grade National Institutes of Health Stroke Scale scores with an accuracy of 83.3% and an AUC of 0.912.

Artificial intelligence models have also been employed to differentiate between ischemic stroke and stroke mimics, such as using artificial neural networks in the ED to recognize acute ischemic stroke. ¹⁰⁹ Similarly, ML models using National Institutes of Health Stroke Scale components have been successful in predicting LVOs in the ED. ¹⁰⁰

Although progress has been made in applying artificial intelligence methods for prehospital stroke management, challenges include attaining sufficient specificity and sensitivity, determining the best strategies to link automated stroke symptom detection to automated notification for help and validating findings in external and larger patient cohorts.

CONCLUSIONS

This systematic review on non‐CT‐based technologies highlights a broad spectrum of prehospital tools for stroke diagnosis. Although these tools aim to facilitate early diagnosis and streamline the triage process, a critical analysis indicates that many technologies may provide results only by complimentary use, suggesting a multimodal approach to optimizing prehospital stroke care. Although we conducted a comprehensive review of the technologies, their heterogeneity precluded meta‐analysis of the data.

MWI offers a portable, possibly low‐cost method for stroke diagnosis and the potential of distinguishing between ischemic and hemorrhagic strokes. Early studies show promise, but further research is needed to validate its accuracy for prehospital use. Portable EEG devices have shown moderate to high accuracy in prehospital LVO detection. Quick application times and tolerability are clear advantages, but limitations include high dropout rates due to bad signal quality. The very limited capability to distinguish ischemic from hemorrhagic strokes precludes application to the use of recombinant tissue plasminogen activator in prehospital settings. NIRS potentially offers a straightforward approach for detecting superficial cranial hemorrhages, though its applicability for deeper hemorrhages and ischemic strokes is constrained. NIRS's limited specificity could be mitigated by combining it with other biomarkers or ultrasound, potentially allowing a layered approach to hemorrhage detection. A downside of many of the discussed technologies is their inability to produce visual images; radiologists and neurologists are accustomed to making diagnostic decisions based on actual imaging rather than on numerical data alone.

Portable MRI, in contrast, shows potential for stroke imaging, allowing bedside examinations in critically ill patients. However, long image acquisition times can delay treatment, and current systems for potential prehospital use lack essential capabilities like perfusion imaging and time‐of‐flight angiography. Additionally, concerns about motion artifacts and modest quality of diffusion‐weighted imaging limit its diagnostic applicability.

Transcranial ultrasound holds significant promise particularly for LVO detection, although examiner dependence is a limitation. Telemedicine or robotic devices may alleviate this issue. Further research is needed on hemorrhage detection, but ultrasound's portability, low cost, and speed make it a strong candidate for broad use. Combining it with NIRS could potentially mitigate its weaknesses in superficial hemorrhage detection and offer almost complete coverage for ICH exclusion. ¹¹⁰

ECD is still in an experimental phase but shows promise as a portable, low‐cost and safe method to localize and characterize stroke lesions. Current data are insufficient to evaluate diagnostic accuracy and further research is therefore needed to determine ECD's effectiveness across stroke types, sizes, and locations and to assess the impact of operator skills on scan quality.

Although only a single study has evaluated a volumetric impedance phase shift spectroscopy device, its portability and high accuracy in detecting severe strokes make it a promising tool for prehospital triage.

Telemedicine has proven to be effective for remote stroke assessment, with high diagnostic accuracy, improved process times, and growing reliability due to advancements in telecommunication. The continued enhancement of connectivity enables remote use of new diagnostic modalities with high bandwidth requirements. Combined with robotic technologies for automated examinations, it might allow for widespread use of specialized assessments like transcranial color‐coded duplex sonography.

New technologies will provide an added value in prehospital stroke management only if they exceed the accuracy of prehospital stroke scales such as the Rapid Arterial Occlusion Evaluation E or the Los Angeles Motor Scale scale–either alone or in conjunction with such scales. This has been shown for telemedicine and in a small study for volumetric impedance phase shift spectroscopy but not yet for other technologies, although biomarkers such as GFAP can obviously identify intracerebral hemorrhages better than clinical assessments.

Future research should prioritize the clinical integration of these tools, especially by leveraging advances in ML, deep learning, and telecommunication technologies. Before new technologies are implemented in routine emergency care, they need to be thoroughly tested in the prehospital setting, including evaluations of their effects on process quality (eg, times from EMS arrival to start of treatment) and clinical outcomes. These innovations may significantly affect stroke survival and recovery by accelerating prehospital stroke diagnosis and enabling targeted, timely treatment.

Sources of Funding

This review did not receive any specific grant from funding agencies in the public, commercial, or not‐for‐profit sectors.

Disclosures

Heinrich J. Audebert declares having received consultancy and speaker honoraria from Bayer Healthcare, Boehringer Ingelheim, Bristol Myers Squibb, Novo Nordisk, Eli Lilly and Sanofi. He has received institutional grants from the German Research Foundation, the German Federal Ministry of Education and research, the Stiftung Deutsche Schlaganfall‐Hilfe, and the German Innovation Fund. During the preparation of the current work, he served as president of the PREhospital Stroke Organisation. Joan Montaner holds the patent “Method for selecting a patient for a reperfusion therapy” EP19382384.6 and is cofounder and shareholder of ABCDx. Sean I. Savitz has served in the following roles as an employee of the institution UTHealth: as a site investigator in clinical trials sponsored by industry companies Brain.Q, IPSEN, Neurastasis, for which UTHealth receives payments on the basis of clinical trial contracts; as an investigator on clinical trials or supported by National Institutes of Health grants; as an investigator in the National Institutes of Health funded Clinical and Translational Science Awards Program; as a principal investigator or coinvestigator on National Institutes of Health‐funded grants in basic science and clinical research, as a principal investigator in state funded projects under the auspices of the LoneStar Stroke Consortium. In his capacity as a UTHealth employee, Sean I. Savitz has provided consulting services on behalf of UTHealth to Lumosa, Neurexcell, Abbvie, Neurastasis, ArunA, Ceribell, AI‐Stroke, CerebralAI, Bayer, CellProthera, DE Shaw, APT Therapeutics, and Pikralida. All compensation from such consulting arrangements has been paid to UTHealth. Guillaume Turc declares having received lecture fees from Guerbet France, consultancy fees from NeuroLogica Corporation and AI‐Stroke. All other authors declare no relevant conflicts of interest related to this work.

Supporting information

Data S1. Supplemental Material

PRISMA 2020 Checklist