PupilScreen

Smartphone pupillometry with machine learning differentiates ischemic from hemorrhagic stroke: A pilot study

OBJECTIVES

Similarities between acute ischemic and hemorrhagic stroke make diagnosis and triage challenging. We studied a smartphone-based quantitative pupillometer for differentiation of acute ischemic and hemorrhagic stroke.

MATERIALS AND METHODS

Stroke patients were recruited prior to surgical or interventional treatment. Smartphone pupillometry was used to quantify components of the pupillary light reflex (PLR). A synthetic minority oversampling technique (SMOTE) was applied to correct sample size imbalance. Four binary classification model types were trained using all possible combinations of the PLR components with 10-fold cross validation stratified by cohort. Models were evaluated for accuracy, sensitivity, specificity, area under the curve (AUC), and F1 score. The three best-performing models were selected based on AUC. Shapley additive explanation plots were produced to explain PLR parameter impacts on model predictions.

RESULTS

Eleven subjects with intraparenchymal hemorrhage and 22 subjects with acute ischemic stroke were enrolled. One way ANOVA demonstrated significant differences between healthy control data, AIS, and IPH in five out of seven PLR parameters. After SMOTE, each class had n=22 PLR recordings for model training. The best-performing model was random forest using a combination of latency, mean and maximum constriction velocity, and mean dilation velocity to discriminate between stroke types with 91.5% (95% confidence interval: 84.1-98.9) accuracy, 90% (82.9-97.1) sensitivity, 93.3% (83-100) specificity, 0.917 (0.847-0.987) AUC, and 90.7% (84.1-97.3) F1 score.

CONCLUSIONS

Smartphone-based quantitative pupillometry could be useful in differentiating between acute ischemic and hemorrhagic stroke.

Comparison of an AI-based mobile pupillometry system and NPi-200 for pupillary light reflex and correlation with glaucoma-related markers

Introduction

Glaucoma is a leading cause of blindness, often progressing asymptomatically until significant vision loss occurs. Early detection is crucial for preventing irreversible damage. The pupillary light reflex (PLR) has proven useful in glaucoma diagnosis, and mobile technologies like the AI-based smartphone pupillometer (AI Pupillometer) offer a promising solution for accessible screening. This study assesses the reliability of the AI Pupillometer in detecting glaucoma.

Methods

In Experiment 1, 20 healthy participants were assessed using both the AI Pupillometer and the NPi-200 device to evaluate equivalence in measuring PLR. Each eye underwent three trials. Experiment 2 included 46 participants, 24 with primary open-angle glaucoma (POAG) and 22 healthy controls. PLR measurements from the AI Pupillometer were correlated with structural and functional ocular parameters. An additional study expanded the sample to 387 participants (103 glaucoma patients, 284 controls), focusing on differential pupillometry parameters to minimize ambient light interference.

Results

In Experiment 1, the AI Pupillometer demonstrated strong correlations with the NPi-200 in key parameters like initial pupil size (r = 0.700), constricted pupil size (r = 0.755), and constriction velocity (r = 0.541), confirming its reliability. In Experiment 2, although no statistically significant differences in light-corrected PLR parameters were found between groups, glaucoma patients had a marginally higher constricted pupil size (p = 0.1632). Significant correlations were observed between pupillometry and advanced ocular imaging results, notably between constriction amplitude and visual field loss. The additional study revealed significant differences in constriction amplitude (p = 0.014) and relative pupil size change (p = 0.0072) between glaucoma patients and controls, reinforcing the AI Pupillometer's diagnostic potential.

Conclusion

This study confirms the AI Pupillometer as a reliable, accessible tool for glaucoma screening. Mobile diagnostics could enhance early detection, improving outcomes for glaucoma patients.

Smartphone-Based Pupillometry Using Machine Learning for the Diagnosis of Sports-Related Concussion

Background: Quantitative pupillometry has been proposed as an objective means to diagnose acute sports-related concussion (SRC). Objective: To assess the diagnostic accuracy of a smartphone-based quantitative pupillometer in the acute diagnosis of SRC. Methods: Division I college football players had baseline pupillometry including pupillary light reflex (PLR) parameters of maximum resting diameter, minimum diameter after light stimulus, percent change in pupil diameter, latency of pupil constriction onset, mean constriction velocity, maximum constriction velocity, and mean dilation velocity using a smartphone-based app. When an SRC occurred, athletes had the smartphone pupillometry repeated as part of their concussion testing. All combinations of the seven PLR parameters were tested in machine learning binary classification models to determine the optimal combination for differentiating between non-concussed and concussed athletes. Results: 93 football athletes underwent baseline pupillometry testing. Among these athletes, 11 suffered future SRC and had pupillometry recordings repeated at the time of diagnosis. In the machine learning pupillometry analysis that used the synthetic minority oversampling technique to account for the significant class imbalance in our dataset, the best-performing model was a random forest algorithm with the combination of latency, maximum diameter, minimum diameter, mean constriction velocity, and maximum constriction velocity PLR parameters as feature inputs. This model produced 91% overall accuracy, 98% sensitivity, 84.2% specificity, area under the curve (AUC) of 0.91, and an F1 score of 91.6% in differentiating between baseline and SRC recordings. In the machine learning analysis prior to oversampling of our imbalanced dataset, the best-performing model was k-nearest neighbors using latency, maximum diameter, maximum constriction velocity, and mean dilation velocity to produce 82% accuracy, 40% sensitivity, 87% specificity, AUC of 0.64, and F1 score of 24%. Conclusions: Smartphone pupillometry in combination with machine learning may provide fast and objective SRC diagnosis in football athletes.

Smartphone pupillometry for detection of cerebral vasospasm after aneurysmal subarachnoid hemorrhage

OBJECTIVES

Vasospasm is a complication of aneurysmal subarachnoid hemorrhage (aSAH) that can change the trajectory of recovery and is associated with morbidity and mortality. Earlier detection of vasospasm could improve patient outcomes. Our objective is to evaluate the accuracy of smartphone-based quantitative pupillometry in the detection of radiographic vasospasm and delayed cerebral ischemia (DCI) after aSAH.

MATERIALS AND METHODS

We prospectively collected pupillary light reflex (PLR) parameters from patients with aSAH admitted to a neurocritical care unit at a single hospital twice daily using quantitative smartphone pupillometry recordings. PLR parameters included: Maximum pupil diameter, minimum pupil diameter, percent change in pupil diameter, latency in beginning of pupil constriction to light, mean constriction velocity, maximum constriction velocity, and mean dilation velocity. Two-tailed t-tests for independent samples were performed to determine changes in average concurrent PLR parameter values between the following comparisons: (1) patients with and without radiographic vasospasm (defined by angiography with the need for endovascular intervention) and (2) patients with and without DCI.

RESULTS

49 subjects with aSAH underwent 323 total PLR recordings. For PLR recordings taken with (n=35) and without (n=241) radiographic vasospasm, significant differences were observed in MIN (35.0 ± 7.5 pixels with vasospasm versus 31.6 ± 6.2 pixels without; p=0.002). For PLR recordings taken with (n=43) and without (n=241) DCI, significant differences were observed in MAX (48.9 ± 14.3 pixels with DCI versus 42.5 ± 9.2 pixels without; p<0.001).

CONCLUSIONS

Quantitative smartphone pupillometry has the potential to be used to detect radiographic vasospasm and DCI after aSAH.

Development of a Smartphone-Based System for Intrinsically Photosensitive Retinal Ganglion Cells Targeted Chromatic Pupillometry

Chromatic Pupillometry, used to assess Pupil Light Reflex (PLR) to a coloured light stimulus, has regained interest since the discovery of melanopsin in the intrinsically photosensitive Retinal Ganglion Cells (ipRGCs). This technique has shown the potential to be used as a screening tool for neuro-ophthalmological diseases; however, most of the pupillometers available are expensive and not portable, making it harder for them to be used as a widespread screening tool. In this study, we developed a smartphone-based system for chromatic pupillometry that allows targeted stimulation of the ipRGCs. Using a smartphone, this system is portable and accessible and takes advantage of the location of the ipRGCs in the perifovea. The system incorporates a 3D-printed support for the smartphone and an illumination system. Preliminary tests were carried out on a single individual and then validated on eleven healthy individuals with two different LED intensities. The average Post-Illumination Pupil Light Response 6 s after the stimuli offsets (PIPR-6s) showed a difference between the blue and the red stimuli of 9.5% for both intensities, which aligns with the studies using full-field stimulators. The results validated this system for a targeted stimulation of the ipRGCs for chromatic pupillometry, with the potential to be a portable and accessible screening tool for neuro-ophthalmological diseases.

A smartphone pupillometry tool for detection of acute large vessel occlusion

OBJECTIVES

Pupillary light reflex (PLR) parameters can be used as quantitative biomarkers of neurological function. Since digital infrared pupillometry is expensive, we sought to examine alterations in PLR parameters using a smartphone quantitative pupillometry platform in subjects with acute ischemic stroke (AIS).

MATERIALS AND METHODS

Patients were enrolled if they presented to the emergency department as a stroke code activation and had evidence of a large vessel occlusion (LVO) on computed tomography angiography. Controls were enrolled from hospital staff. A smartphone pupillometer was used in AIS patients with LVO pre-mechanical thrombectomy, immediately post-thrombectomy, and at 24 h post-thrombectomy. Clinical and demographic data were collected, along with the proprietary Neurological Pupil index (NPi) score from the NPi-200 digital infrared pupillometer. PLR parameters were compared using mean differences. The absolute and non-absolute inter-eye difference in each parameter for each subject were also analyzed by measuring 1 - (R:L) to determine alteration in the equilibrium between subject pupils. The NPi was analyzed for mean differences between cohorts.

RESULTS

Healthy controls (n = 132) and AIS patients (n = 31) were enrolled. Significant differences were observed in PLR parameters for healthy subjects when compared to pre-thrombectomy subjects in both mean and absolute inter-eye differences after post hoc Bonferroni correction. The proprietary NPi score was not significantly different for all groups and comparisons.

CONCLUSIONS

Significant alterations in the PLR were observed in AIS patients with LVO before thrombectomy, indicating the potential use of smartphone pupillometry for detection of LVO.

Racially fair pupillometry measurements for RGB smartphone cameras using the far red spectrum

Pupillometry is a measurement of pupil dilation commonly performed as part of neurological assessments. Prior work have demonstrated the potential for pupillometry in screening or diagnosing a number of neurological disorders including Alzheimer's Disease, Schizophrenia, and Traumatic Brain Injury. Unfortunately, the expense and inaccessibility of specialized pupilometers that image in the near infrared spectrum limit the measurement to high resource clinics or institutions. Ideally, this measurement could be available via ubiquitous devices like smartphones or tablets with integrated visible spectrum imaging systems. In the visible spectrum of RGB cameras, the melanin in the iris absorbs light such that it is difficult to distinguish the pupil aperature that appears black. In this paper, we propose a novel pupillometry technique to enable smartphone RGB cameras to effectively differentiate the pupil from the iris. The proposed system utilizes a 630 nm long-pass filter to image in the far red (630-700 nm) spectrum, where the melanin in the iris reflects light to appear brighter in constrast to the dark pupil. Using a convolutional neural network, the proposed system measures pupil diameter as it dynamically changes in a frame by frame video. Comparing across 4 different smartphone models, the pupil-iris contrast of N = 12 participants increases by an average of 451% with the proposed system. In a validation study of N = 11 participants comparing the relative pupil change in the proposed system to a Neuroptics PLR-3000 Pupillometer during a pupillary light response test, the prototype system acheived a mean absolute error of 2.4%.

At-Home Pupillometry using Smartphone Facial Identification Cameras

With recent developments in medical and psychiatric research surrounding pupillary response, cheap and accessible pupillometers could enable medical benefits from early neurological disease detection to measurements of cognitive load. In this paper, we introduce a novel smartphone-based pupillometer to allow for future development in clinical research surrounding at-home pupil measurements. Our solution utilizes a NIR front-facing camera for facial recognition paired with the RGB selfie camera to perform tracking of absolute pupil dilation with sub-millimeter accuracy. In comparison to a gold standard pupillometer during a pupillary light reflex test, the smartphone-based system achieves a median MAE of 0.27mm for absolute pupil dilation tracking and a median error of 3.52% for pupil dilation change tracking. Additionally, we remotely deployed the system to older adults as part of a usability study that demonstrates promise for future smartphone deployments to remotely collect data in older, inexperienced adult users operating the system themselves.

Pupillary light reflex as a diagnostic aid from computational viewpoint: A systematic literature review

This work presents a detailed and complete review of publications on pupillary light reflex (PLR) used to aid diagnoses. These are computational techniques used in the evaluation of pupillometry, as well as their application in computer-aided diagnoses (CAD) of pathologies or physiological conditions that can be studied by observing the movements of miosis and mydriasis of the human pupil. A careful survey was carried out of all studies published over the last 10 years which investigated, electronic devices, recording protocols, image treatment, computational algorithms and the pathologies related to PLR. We present the frontier of existing knowledge regarding methods and techniques used in this field of knowledge, which has been expanding due to the possibility of performing diagnoses with high precision, at a low cost and with a non-invasive method.

Artificial Intelligence Segmented Dynamic Video Images for Continuity Analysis in the Detection of Severe Cardiovascular Disease

In this paper, an artificial intelligence segmented dynamic video image based on the process of intensive cardiovascular and cerebrovascular disease monitoring is deeply investigated, and a sparse automatic coding deep neural network with a four layers stack structure is designed to automatically extract the deep features of the segmented dynamic video image shot, and six categories of normal, atrial premature, ventricular premature, right bundle branch block, left bundle branch block, and pacing are achieved through hierarchical training and optimization. Accurate recognition of heartbeats with an average accuracy of 99.5%. It provides technical assistance for the intelligent prediction of high-risk cardiovascular diseases like ventricular fibrillation. An intelligent prediction algorithm for sudden cardiac death based on the echolocation network was proposed. By designing an echolocation network with a multilayer serial structure, an intelligent distinction between sudden cardiac death signal and non-sudden death signal was realized, and the signal was predicted 5 min before sudden death occurred, with an average prediction accuracy of 94.32%. Using the self-learning capability of stack sparse auto-coding network, a large amount of label-free data is designed to train the stack sparse auto-coding deep neural network to automatically extract deep representations of plaque features. A small amount of labeled data then introduced to micro-train the entire network. Through the automatic analysis of the fiber cap thickness in the plaques, the automatic identification of thin fiber cap-like vulnerable plaques was achieved, and the average overlap of vulnerable regions reached 87%. The overall time for the automatic plaque and vulnerable plaque recognition algorithm was 0.54 s. It provides theoretical support for accurate diagnosis and endogenous analysis of high-risk cardiovascular diseases.

Guidelines and Benchmarks for Deployment of Deep Learning Models on Smartphones as Real-Time Apps

Deep learning solutions are being increasingly used in mobile applications. Although there are many open-source software tools for the development of deep learning solutions, there are no guidelines in one place in a unified manner for using these tools toward real-time deployment of these solutions on smartphones. From the variety of available deep learning tools, the most suited ones are used in this paper to enable real-time deployment of deep learning inference networks on smartphones. A uniform flow of implementation is devised for both Android and iOS smartphones. The advantage of using multi-threading to achieve or improve real-time throughputs is also showcased. A benchmarking framework consisting of accuracy, CPU/GPU consumption, and real-time throughput is considered for validation purposes. The developed deployment approach allows deep learning models to be turned into real-time smartphone apps with ease based on publicly available deep learning and smartphone software tools. This approach is applied to six popular or representative convolutional neural network models, and the validation results based on the benchmarking metrics are reported.