Prediction of anemia in real-time using a smartphone camera

RESEARCH ARTICLE

Prediction of anemia in real-time using a

smartphone camera processing conjunctival

images

Leon Zhao

1☯

, Alisa Vidwans

, Courtney J. Bearnot

, James Rayner

, Timmy Lin

Janette Baird

, Selim Suner

, Gregory D. Jay

1☯

1 The Warren Alpert Medical School, Brown University, Providence, Rhode Island, United States of America,

2 Rhode Island Hospital, Providence, Rhode Island, United States of America

☯ These authors contributed equally to this work.

* [email protected]

Abstract

Anemia is defined as a low hemoglobin (Hb) concentration and is highly prevalent world-

wide. We report on the performance of a smartphone application (app) that records images

in RAW format of the palpebral conjunctivae and estimates Hb concentration by relying

upon computation of the tissue surface high hue ratio. Images of bilateral conjunctivae were

obtained prospectively from a convenience sample of 435 Emergency Department patients

using a dedicated smartphone. A previous computer-based and validated derivation data

set associating estimated conjunctival Hb (HBc) and the actual laboratory-determined Hb

(HBl) was used in deriving Hb estimations using a self-contained mobile app. Accuracy of

HBc was 75.4% (95% CI 71.3, 79.4%) for all categories of anemia, and Bland-Altman plot

analysis showed a bias of 0.10 and limits of agreement (LOA) of (-4.73, 4.93 g/dL). Analysis

of HBc estimation accuracy around different anemia thresholds showed that AUC was maxi-

mized at transfusion thresholds of 7 and 9 g/dL which showed AUC values of 0.92 and 0.90

respectively. We found that the app is sufficiently accurate for detecting severe anemia and

shows promise as a population-sourced screening platform or as a non-invasive point-of-

care anemia classifier.

Introduction

Anemia, defined as a low blood hemoglobin (Hb) concentration, has a global prevalence esti-

mated at 22.8%, or 50.3 million years lived with disability in 2019 [1]. Young children, men-

struating women, and pregnant and postpartum women are particularly affected by iron

deficiency anemia [2]. The World Health Organization (WHO) estimates that 40% of children

under the age of five in 35 countries have anemia and has classified anemia as a severe public

health problem [3]. Anemia can result in a range of clinical complications, such as immune

system dysfunction, impaired thermoregulation and neurocognitive function, and slow cogni-

tive and motor development. Understanding the socioeconomic determinants and developing

new tools for early diagnosis should remain a priority.

PLOS ONE

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 1 / 13

a1111111111

OPEN ACCESS

Citation: Zhao L, Vidwans A, Bearnot CJ, Rayner J,

Lin T, Baird J, et al. (2024) Prediction of anemia in

real-time using a smartphone camera processing

conjunctival images. PLoS ONE 19(5): e0302883.

https://doi.org/10.1371/journal.pone.0302883

Editor: Germana Bancone, Shoklo Malaria

Research Unit, THAILAND

Received: September 19, 2023

Accepted: April 15, 2024

Published: May 13, 2024

access article distributed under the terms of the

Creative Commons Attribution License, which

permits unrestricted use, distribution, and

reproduction in any medium, provided the original

author and source are credited.

Data Availability Statement: https://zenodo.org/

record/8277462.

Funding: Brown Physicians Inc. Grant #: 7132042

awarded to SS and JR. The funding sources had no

role in the study design, in data collection, analysis,

or interpretation, in manuscript writing, or in the

decision to submit the study results for peer-

reviewed publication.

Competing interests: I have read the journal’s

policy and the authors of this manuscript have the

following competing interests: GD Jay and S.

Suner authored US Patent #7,711,403. GD Jay, S.

The standard test for diagnosing anemia is the Complete Blood Count (CBC). While a CBC

measures the severity of anemia, it requires venipuncture, trained phlebotomists, laboratory

technicians, the use of chemical reagents, and dedicated lab equipment [4]. In resource-rich

hospitals, it takes 1–4 hours to obtain results. Therefore, the collection of CBC is often limited

to geographical regions with adequate healthcare infrastructure. However, studies have shown

that anemia is disproportionally prevalent in rural areas that lack these resources and visualiza-

tion of conjunctival pallor is relied upon to qualitatively support anemia but is unreliable [5,

6]. As a result, there is an unmet need for accessible, non-invasive, point-of-care (POC) tools

that can screen for anemia rapidly. Non-invasive measurement of Hb [4, 7–10] has been made

possible through smartphone imaging of the palpebral conjunctiva [4, 7–10]. The light

reflected off the palpebral conjunctiva membrane can be evaluated by digital photographs, and

studies have shown that images in RAW [4] and RGB format [7–10] obtained from a smart-

phone camera can predict Hb concentration, albeit imprecisely. Apps that process nail-bed

images are similarly imprecise upon real-world testing [11, 12]. Lack of precision using a

smartphone app is outweighed by the sheer accessibility of smartphones to the general popula-

tion. The ability to assess more patients through the ubiquity of cell phones presents a near-

term opportunity for universal screening for anemia through telehealth apps.

Studies of POC devices for non-invasive Hb prediction appeared promising based on initial

foundational data sets but were not validated in many instances [7, 13–15] which are summa-

rized in S1 Table. Measuring Hb concentration post-hoc from palpebral conjunctiva images

has a reported accuracy of 82.9% in an adult population [4]. Other studies using cell phone

enabled apps reporting Hb in real time [8, 9, 16] report similar results (S1 Table). Given the

previous limitations in accuracy, our study aimed to evaluate a self-contained app and user

interface as a tool for anemia severity classification in real-time using RAW images processed

in a 32 bit enabled classifier. The novelty and innovation of the present work is to show that a

self-contained POC device can estimate Hb concentration around critical and actionable

transfusion thresholds by imaging an area that physicians are already familiar with- the con-

junctiva. Determining the urgent need for blood transfusion is a suitable gateway for early

adoption of this computerized app by emergency and intensive care medical personnel. In

these situations, the hypothetical risks of empirical measurement may be counterweighted by

the risk of failing to appreciate patients’ severe anemia. The app is also user friendly in so far

that self-measurement is possible.

Materials and methods

Study design

This was a prospective observational validation study of a convenience sample of adult patients

presenting to the Emergency Department (ED) between June 18, 2022 and February 10, 2023.

Patients over the age of 18 years presenting with any chief complaint and having 1) a CBC

obtained as part of their medical care, 2) the ability to provide informed consent, and 3) the

ability to expose the palpebral conjunctiva of both eyes, were approached for participation.

Patients with injury, jaundice, inflammation, or infection of either eye were excluded. This

study followed the Standards for Reporting of Diagnostic Accuracy guidelines (STARD) 2015

[17], which provides a checklist of essential items for reporting diagnostic accuracy studies.

Study population

Lifespan Institutional Review Board approved this study (Ref # 209416). All patients who par-

ticipated provided written informed consent. The individual pictured in Fig 1 has provided

written informed consent (as outlined in PLOS consent form) to publish their image alongside

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 2 / 13

Suner and J. Rayner have submitted US Patent

application #17/789,740. This does not alter our

adherence to PLOS ONE policies on sharing data

and materials.

the manuscript. Participating patients were identified based on a Best Practice Alert (BPA) in

the EPIC electronic medical record that notified research assistants (RA) when a patient had

an abnormal Hb level as determined by their CBC. Abnormal was defined as <13.5 g/dL or

>16.0 g/dL for men and <12.0 g/dL or >15.0 g/dL for women. Research assistants consented

and enrolled participants, and then acquired conjunctival images within 4 hours of venipunc-

ture for CBC determination.

Demographic information was collected along with hospital laboratory-reported test results

including blood oxygen saturation, HBl (laboratory-determined Hb), and time of collection. The

specific causes of anemia were not a subject of study and thus were not included in the database.

Anemia classification binning

Five anemia classifications were identified: mild anemia (� 10 and < 12 g/dL) for women,

mild anemia (� 10 and < 13.5 g/dL) for men [18], moderate (� 7 and < 10 g/dL) [19], severe

(� 5 and < 7 g/dL) [16] and extreme (< 5 g/dL) for men and women [20, 21]. The classifica-

tion bin for mild anemia also included the participants who were normemic (�12 g/dL) for

women and (� 13.5 g/dL) for men [18].

Fig 1. Smartphone app interface and camera positioning. Demonstration of use of an eMoglobin app enabled smart phone on a volunteer

member of the research team, A) 3–6 inches away depending upon ambient lighting and image clarity (B) representative captured images and

HBc estimated Hb values.

https://doi.org/10.1371/journal.pone.0302883.g001

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 3 / 13

Image capture and processing

Participants were asked to remain still and retract their lower eyelids to expose the conjunctiva

with the assistance of research assistants (RA) if necessary. Images were obtained in the eMo-

globin app on an iPhone X with iOS version 15.6.1 (Apple Inc, Cupertino, CA) under ambient

indoor light using the back camera held approximately three to six inches from the target con-

junctiva (Fig 1). Tapping the screen to focus on the conjunctiva activated the auto-focus and

brightness adjustment of the smartphone camera. A minimum of four images were obtained

from each patient: two images of the left and right eye. Images were recorded in RAW format

without flash and images were encoded at a 32-bit level of resolution (S1 Fig) [4]. After each

image was taken, the app displayed the image for the research assistant to visually inspect and

select the region of interest (ROI). The research staff were instructed to adjust the ROI so that

the displayed image was centered on the patient’s palpebral conjunctiva. Selected ROI images

were stored and processed as described previously [4] and analyzed in the eMoglobin app in

real time. High hue ratio was calculated from each image and correlated to the HBl from the

previously validated correlation. Conjunctival Hb concentration estimates (HBc) were avail-

able instantly [4].

Statistical methods

Statistical analysis was conducted using SAS (version 9.4, Cary, NC). Descriptive statistics

were utilized to characterize the study population by age, race, and skin color [4] and describe

the distribution of HBl values. Polynomial regression was performed to assess the association

between HBl and HBc values, including a quadratic term for HBc. A regression analysis was

also performed to assess the relationship between high hue ratio and HBl. Additionally, a

mixed effects regression analysis was conducted with RA as the random effects to assess the

variance in the association between HBc and HBl accounted for by RA operator performance

(n = 16). Due to variability in the numbers of patients assessed by the RA operators, we catego-

rized this into a 2-level of assessed volume-based categories for RA’s (< 20, �20 patients). The

intraclass correlation coefficient (ICC) was calculated to assess the operator effect.

Clinical usefulness of HBc

HBl was categorized as anemic (<12 g/dL for women and < 13.5 g/dL for men) or not anemic

as defined by the American Society of Hematology (ASH) [22]. The association between HBc

and HBl across different levels of anemia described above was evaluated using 2x2 tables to cal-

culate accuracy, sensitivity, specificity, false positive and false negative rates. Additionally, dis-

crimination was evaluated by the Area Under Curve (AUC) of the receiver operating

characteristic (ROC) curves using logistic regression.

Bland-Altman plots

Bland-Altman plots assessed the agreement between HBc and HBl measures. The Bland-Alt-

man plot depicts the difference between HBc and HBl plotted against the average of HBc and

HBl measures.

Results

During the enrollment period between June 18, 2022, and February 10, 2023, images from 435

unique participating patients were obtained. Data from nine subjects were excluded because

each had only one image recorded, leaving 426 subjects for the final analysis. The average age

was 53 years (range 19–94) and 55.2% were women. The study population was roughly

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 4 / 13

representative of the racial and ethnic distribution of Rhode Island in the 2020 census [23] -

72.8% White and 12.7% Hispanic or Latinx (S2 Table). The predominant skin color was light

(75.6%), 424 patients (99.6%) did not have jaundice, and 413 (97%) had normal oxygen satura-

tion levels (mean oxygen saturation: 97.46%; 95% CI 97.44, 97.47%). Mean HBl was 12.6 g/dL

(range 3.2–18.8 g/dL, 95% CI 6.8, 18.3 g/dL) and the distribution of HBl is illustrated in S2 Fig.

The average of the four app-derived Hb values was calculated to produce a HBc estimate.

HBc and HBl were plotted and a visual inspection of the best line of fit and residuals (Fig 2),

indicated that a quadratic fit best identified the relationship between HBc and HBl. The inter-

cept was = 0 (95% CI -0.072, 0.072), and the quadratic of HBc was 0.66 (95% CI 0.59, 0.73). A

slope = 1.0 indicated perfect fit. This suggests that HBc underestimates HBl.

The majority (81.7%) of participants had mild anemia by serology (Table 1A). There were

comparatively fewer subjects with severe (3.8%) and extreme (1.6%) anemia respectively. The

app was most accurate in the extreme anemia category, with agreement between HBc and HBl

being 100% (Table 1B). The app colocalized HBl and HBc estimates in 25.3% of the mildly ane-

mic patients. The HBc estimates incorrectly categorized normemic subjects of both sexes as

mildly anemic in 60 of the 95 mildly anemic patients (63.2%). Predictions in the severe (�5

and < 7 g/dL) and moderate (� 7 and < 10 g/dL) categories had greater variability, with agree-

ment of 18.2% and 25.0%, respectively (Table 1B).

Fig 2. Correlation between HBc and HBl. Correlation between conjunctiva-estimated Hb (HBc) (average of all four replicates) vs actual

measured Hb (HBl) in pre-defined operating range (2–14 g/dl). There was a moderate degree of association (p < .001) between HBc and

HBl, with correlation coefficients of 0.65 and 0.66 for HBc and the quadratic of HBc respectively. These correlation coefficients do not

represent the actual association between HBc and HBl as HBc prediction values were limited to 14 g/dL in the app display, contributing to

clustering at HBc at 14 g/dL due to the processing and display limitation of the app which does not execute calculations and prediction is

excess of 14 g/dL.

https://doi.org/10.1371/journal.pone.0302883.g002

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 5 / 13

Accuracy of HBc was 75.4% (95% CI 71.3, 79.4) for all categories of anemia, and was signifi-

cantly lower for women (AUC = 0.74) compared to men (AUC = 0.79). Sensitivity and speci-

ficity of HBc for predicting anemia in men and women combined was 54.3 (95% CI 46.9, 61.8)

and 89.7 (95% CI 86.0, 93.5), respectively. Accuracy, sensitivity, and specificity of HBc for pre-

dicting anemia at a high transfusion threshold (<9 g/dL) were greater with values of 91.1 (95%

CI 88.4, 93.8), 58.3 (95% CI 45.9, 70.8), and 96.5 (95% CI 94.6, 98.3), respectively. Accuracy,

sensitivity, and specificity of HBc for predicting anemia at a low transfusion threshold (<7 g/

dL) were greater with values of 92.7 (95% CI 90.3, 95.2), 69.6 (95% CI 50.8, 88.4), and 94.0

(95% CI 91.7, 96.4), respectively (S3 Table).

ROC curves were generated to analyze the effect of increasing transfusion threshold on the

sensitivity and specificity of HBc (Fig 3). AUC overall for sex-specific cut-off was 0.76 (95% CI

0.72, 0.81) and the AUC was not significantly different between women and men cut-off values

(women = 0.74 (95% CI 0.68,0.81); men = 0.79 (95% CI 0.73,0.85)). AUC for the restrictive

transfusion threshold was 0.92 (95% CI 0.85, 0.98), while AUC for the liberal transfusion

threshold was 0.90 (95% CI 0.85, 0.94). The present eMoglobin algorithm had greater diagnos-

tic accuracy for Hb values below the restrictive and liberal transfusion thresholds compared to

values below the threshold values for anemia for men and women.

Table 1. Sub-populations of anemia threshold binning. Population frequency of laboratory measured Hb (HBl) values in anemia classifications. B) Agreement of HBc

and HBl in participants across anemia classifications.

Anemia classification (g/dL) N (%)

Normemic (Women: �12; Men: �13.5) 253 (59.4%)

Mild Anemia (Women: �10 and <12; Men: �10 and <13.5) 95 (22.3%)

Moderate Anemia (�7 and <10) 55 (12.9%)

Severe Anemia (�5 and <7) 16 (3.8%)

Extreme Anemia (<5) 7 (1.6%)

HBI (g/dL)

Normemic (�12 for

Women and �13.5

for Men) (n = 253)

Mild

Anemia

(�10 and

<12 for

Women)

and (�10

and <13.5

for Men)

(n = 95)

Moderate Anemia

(�7 and <10) (n = 55)

Severe Anemia

(�5 and <7) (n = 16)

Extreme Anemia

(<5) (n = 7)

Normemic (Women:

�12; Men: �13.5)

(n = 306)

227 (89.7%) 60 (63.2%) 18 (32.7%) 1 (6.25%) 0 (0%)

Mild Anemia (Women:

�10 and <12; Men:

�10 and <13.5))

(n = 57)

18 (7.1%) 24 (25.3%) 10 (18.2%) 5 (31.25%) 0 (0%)

HBc (g/dL) Moderate Anemia (�7

and <10) (n = 23)

5 (2.0%) 7 (7.4%) 10 (18.2%) 1 (6.25%) 0 (0%)

Severe Anemia (�5 and

<7) (n = 12)

1 (0.4%) 2 (2.1%) 5 (9.1%) 4 (25%) 0 (0%)

Extreme Anemia (<5)

(n = 28)

2 (0.8%) 2 (2.1%) 12 (21.8%) 5 (31.25%) 7 (100%)

HBl is laboratory-measured Hb and HBc is conjunctival-estimated Hb

https://doi.org/10.1371/journal.pone.0302883.t001

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 6 / 13

The Bland-Altman plot exhibits clustering along a diagonal line in the higher range of aver-

age HBl (>10 g/dL) due to the algorithm which set the maximum HBc to 14 g/dL in the app.

Bland-Altman plot analysis shows a bias of 0.10 and limits of agreement (LOA) of (-4.73, 4.93

g/dL). Error was found to trend with increasing average Hb values (slope = 0.68 (0.60, 0.75), p

< .001) (Fig 4).

Intraclass correlation calculation between operators showed that using a 2-level category

(< 20 patients, �20), 3.4% of the variance in HBc predicting HBl was accounted for by the RA

performance. This indicates that the predictive accuracy of HBc from the eMoglobin algorithm

was minimally impacted by interoperator variability (S4 Table).

Discussion

This is the first study to evaluate the clinical utility of a smartphone-based app employing a

novel 32 bit image processing method developed to classify anemia severity using digital

images of the conjunctiva processed locally in real time. Anemia is a pressing health concern

Fig 3. Receiver operator characteristics of HBc using HBl across different anemia guidelines. Receiver operator

characteristic curves for A) overall in both sexes using 12 g/dL as the anemia threshold, B) anemia definitions

differentiated by <12 and <13.5 g/dL for women and men respectively, C) restrictive transfusion threshold of 7 g/dL

defined by the American Association of Blood & Biotherapies (AABB), and D) liberal threshold of 9 g/dL defined by

the AABB. The x-axis depicts 1-specificity and the y-axis shows the sensitivity. The red line represents the ROC and the

black line is the no-discrimination line.

https://doi.org/10.1371/journal.pone.0302883.g003

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 7 / 13

worldwide and disproportionately affects people living in under-resourced areas [5]. Early

screening and detection of anemia is crucial for preventing associated sequelae such as delayed

cognitive and motor development in children, impaired immune response, and increased

maternal mortality [24]. As smartphones become ubiquitous, a non-invasive smartphone-

based tool for anemia screening could serve as a pre-clinical and clinical tool that would allow

for timely and inexpensive anemia screening for patients without access to blood testing.

To validate this smartphone app we used a data set that was collected separately from a pre-

viously reported training and validation set [4]. A separate validation study is an important

requirement in developing these devices to avoid data overfitting where noise in the derivation

data set negatively affects the accuracy of the device. Validation studies using a new cohort of

data have been reported in only four studies developing smartphone-based apps for Hb esti-

mation (S1 Table). In the present study, participant data was collected in an ED clinical setting

and test results were compared to blood-drawn standard measurements. Our app relied on the

smartphone camera for in vivo image collection and did not require external smartphone

attachments [10, 25]. This enables anyone with a smartphone and internet connection to use

Fig 4. Bland-Altman plot for HBc compared to HBl. Bland-Altman plot for HBc (average of the four replicated estimates) compared to HBl (laboratory-

determined) for all 426 participants. The average Hb concentration (HBl+HBc/2) in g/dL on the x-axis is plotted against Hb concentration error (HBl–HBc)

in g/dL on the y-axis. The red shaded area represents limits of agreement. The solid red line represents the mean difference between HBl and HBc, +/- 2

standard deviations of the mean of the differences (dashed red line), +/- 3 standard deviations of the mean of the differences (green dotted line), and the

solid blue line represents 0 error.

https://doi.org/10.1371/journal.pone.0302883.g004

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 8 / 13

the app. Our algorithm utilizes pixel extraction instead of analysis of image RGB values. In pre-

vious methods that used RGB value analysis, which a majority of other Hb prediction apps use

[7–10, 13–15, 26–28], each pixel has a red, green, and blue value that contributes to the color

of the image and is limited to one of 256 values [28]. Using techniques described by Sumner

[29], a RAW image processing algorithm that maximized the color resolution of images was

developed, wherein each color level was encoded with 32 bits giving 4.3 trillion values for each

channel and 8x1028 colors allowing for highly accurate color analysis. These data were further

processed into a high hue ratio [14] which appeared to be more closely approximating Hb val-

ues (S3 Fig). Consistent with our previous study using the same algorithm [4], the use of an

artificial light source did not appear to change the quality of images and affect prediction

accuracy.

Due to previously published limitations in the predictive accuracy of smartphone-imaged

conjunctiva for Hb [4, 7–10], we chose to analyze the predicted Hb values in a role as a screen-

ing device for anemia classification. Discretization of a continuous clinical variable into inter-

vals in theory should partition and maintain original patterns and recapitulate how the data

would be interpreted by the end user, and may also minimize spurious data [30]. For patients

with severe anemia, the app had higher accuracy, sensitivity, and specificity compared to

patients with moderate anemia. These findings suggest that the eMoglobin app is more suitable

for detecting severe anemia rather than mild or moderate anemia. This is useful clinically as

detection of severe anemia is more time-sensitive than identifying mild anemia.

Several apps in development appear promising based on imaging of perfused tissue surfaces

but underperformed compared to co-oximetry and invasive methods [31, 32]. Fingernail bed

imaging [16] was based on an imaging algorithm used for drawn blood [11]. However when

applied non-invasively to fingernail surfaces in Bihar, India to predict Hb values in a popula-

tion of school-aged children the app proved to be highly inaccurate without algorithm training

[12]. The benefit of basing Hb prediction on images of the conjunctiva [26] compared to other

vascularized surfaces such as fingernails is that vascular beds are not affected by factors includ-

ing poor perfusion from vasoconstriction and shock, fungal infections, discoloration from

trauma, fingernail polish and staining, or artificial fingernails [16].

The rise of point-of-care digital health technologies has been catalyzed by the SARS-CoV-2

pandemic and the mobile health market size is expected to reach $189 billion by 2025 [33].

However many mobile health smartphone apps are unregulated and pushed to market before

rigorous validation [34, 35].

Numerous thresholds in classifying anemia exist in the literature [18–20, 36–38]. Reexami-

nation of a normal Hb level associated with the overall clinical well-being of patients in low

and middle-income countries (LMIC) results in a level lower than the WHO 12.0 or 13.0 g/dL

criteria for women and men respectively [37]. The National Cancer Institute cites mild anemia

as 10 g/dL to the lower limit of normal [18]. This range-based definition of mild anemia is sup-

ported in both LMIC [21] and high-income countries [38]. Thresholds for severe anemia coin-

cide with thresholds for transfusion which are stratified as either restrictive (7–8 g/dL) or

liberal transfusion (9–10 g/dL) thresholds in moderate to severe anemia often depending upon

accompanying symptoms and co-morbidities [36]. In Kigali, Rwanda for example 5 g/dL is the

recommended threshold below which blood transfusion is recommended [20]. In the UK and

US this therapeutic recommendation is anchored at 7 g/dL [36]. Thus, the differentiated bins

in this investigation were established a priori in an approach to address the myriad of anemia

definition thresholds and associated blood transfusion guidelines from a global perspective.

The app’s LOA is comparable to that of the Radical-7™ pulse co-oximetry device (Masimo,

Irvine, CA) when first reported, and is a FDA-approved POC Hb measuring device (Bias

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 9 / 13

-1.09, LOA -4.20 to 2.02) [39]. The Masimo Pronto pulse co-oximeter has a sensitivity and

specificity of 86% and 81% respectively in measuring anemia at a threshold of 13.5 g/dL [40].

The present eMoglobin algorithm or other modes [31] of perfused tissue reflectance spec-

trophotometry [4] may be useful in the screening for severe anemia in populations or patients

at risk, respectively in-home use/self-care applications. If it were accepted as an inexpensive

screen, we propose that this metric would be abbreviated as SrHb to quickly communicate

‘reflective spectrophotometry’ as the method of operation for Hb estimation and its inherent

limitations. This abbreviation conveys the similar (but unrelated) abbreviation of SpHb which

refers to continuous Hb monitoring provided by pulse co-oximetry.

The results of this study should be interpreted in the context of certain limitations. First,

RAs had to confirm that a patient’s CBC was taken within the four-hour study window. Thus,

RAs were not blinded to the HBl which could have introduced selection bias when considering

the ROI. Secondly, despite tailoring the enrollment strategy to target patients with moderate to

extreme anemic thresholds, only 78 (18.3%) of participants met these thresholds. The study

likely would have benefitted from enrolling more severely anemic patients to increase the sta-

tistical strength of the specificity calculations. Thirdly, lighting was not rigorously controlled

and some variation in Hb estimation is possible. Future work would include efforts to under-

stand intra-subject variation such as studying participants undergoing blood transfusion with

repeated measures.

Conclusion

The findings from this study demonstrate that a smartphone app can serve as a useful tool for

non-invasive severe anemia classification with acceptable accuracy. Given the high prevalence

of anemia and its associated morbidity and mortality, the development of an accurate and

accessible anemia screening tool is essential to reduce the risk of health complications and alle-

viate the social and economic burdens associated with untreated anemia. This app may have

substantial clinical utility in a myriad of settings in which noninvasive anemia screening

would be beneficial.

Supporting information

S1 Fig. Technical description of eMoglobin algorithm.

(DOCX)

S2 Fig. Distribution of laboratory-determined Hb (HBl) values. Subjects’ HBl levels ranged

between 3.2 and 18.8 g/dL (mean 12.6 g/dL) across N = 426 participants.

(TIF)

S3 Fig. Average high hue ratio vs HBl. The high hue ratio (HHR) primary parameter

extracted from the images was the most predictive of HBI. HHR values were correlated with

HBc range of 2 to 14 g/dL The mean HBl intercept was 5.8 (4.9, 6.7) (95% CI; p<0.001). The

mean HHR was 7.6 (6.6, 8.5) (95% CI; p<0.001). Value of mean std error was 5.47 and R-

squared was 0.372.

(DOCX)

S1 Table. Referenced studies of methods to estimate Hb concentration non-invasively by

imaging perfused tissue surfaces.

(DOCX)

S2 Table. Patient demographics.

(DOCX)

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 10 / 13

S3 Table. Clinical usefulness of conjunctiva-estimated Hb (HBc). Accuracy, sensitivity,

specificity, false positive rate, and false negative rate for anemia as defined by the American

Society of Hematology (ASH) and for transfusion thresholds as defined by the Association for

the Advancement of Blood & Biotherapies (AABB) are shown as predicted values [95% CI].

(DOCX)

S4 Table. Intraclass correlation calculation for RAs. Intraclass correlation calculation using

a mixed effect regression analysis predicting laboratory-determined Hb (HBl) measurement

using the eMoglobin data.

(DOCX)

Acknowledgments

We thank the Rhode Island and Miriam Hospital Emergency Department research assistants

for their help in data collection.

Author Contributions

Conceptualization: Selim Suner, Gregory D. Jay.

Data curation: Leon Zhao.

Formal analysis: Leon Zhao, Timmy Lin, Janette Baird.

Investigation: Alisa Vidwans.

Methodology: Courtney J. Bearnot.

Project administration: Leon Zhao, Selim Suner, Gregory D. Jay.

Resources: Selim Suner.

Software: James Rayner.

Supervision: Selim Suner, Gregory D. Jay.

Writing – original draft: Leon Zhao, Alisa Vidwans, Courtney J. Bearnot, Timmy Lin, Janette

Baird, Selim Suner, Gregory D. Jay.

Writing – review & editing: Leon Zhao, Alisa Vidwans, Courtney J. Bearnot, Timmy Lin, Jan-

ette Baird, Selim Suner, Gregory D. Jay.

References

1. Gardner W, Kassebaum N. Global, regional, and national prevalence of anemia and its causes in 204

countries and territories, 1990–2019. Curr Dev Nutr 2020; 4:nzaa053_035. https://doi.org/10.1093/cdn/

nzaa053_035

2. Estimating the burden of disease attributable to iron deficiency anaemia in South Africa in 2000: original

article | South African Medical Journal n.d. https://journals.co.za/doi/abs/10.10520/EJC69061

(accessed May 18, 2023).

3. Sunuwar DR, Singh DR, Pradhan PMS, Shrestha V, Rai P, Shah SK, et al. Factors associated with ane-

mia among children in South and Southeast Asia: a multilevel analysis. BMC Public Health 2023;

23:343. https://doi.org/10.1186/s12889-023-15265-y PMID: 36793012

4. Suner S, Rayner J, Ozturan IU, Hogan G, Meehan CP, Chambers AB, et al. Prediction of anemia and

estimation of hemoglobin concentration using a smartphone camera. PLOS ONE 2021; 16:e0253495.

https://doi.org/10.1371/journal.pone.0253495 PMID: 34260592

5. McLean E, Cogswell M, Egli I, Wojdyla D, De Benoist B. Worldwide prevalence of anaemia, WHO Vita-

min and Mineral Nutrition Information System, 1993–2005. Public Health Nutr 2009; 12:444. https://doi.

org/10.1017/S1368980008002401 PMID: 18498676

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 11 / 13

6. Sheth TN, Choudhry NK, Bowes M, Detsky AS. The relation of conjunctival pallor to the presence of

anemia. J Gen Intern Med 1997; 12:102–6. https://doi.org/10.1046/j.1525-1497.1997.00014.x PMID:

9051559

7. Collings S, Thompson O, Hirst E, Goossens L, George A, Weinkove R. Non-invasive detection of anae-

mia using digital photographs of the conjunctiva. PLOS ONE 2016; 11:e0153286. https://doi.org/10.

1371/journal.pone.0153286 PMID: 27070544

8. Appiahene P, Arthur EJ, Korankye S, Afrifa S, Asare JW, Donkoh ET. Detection of anemia using con-

junctiva images: A smartphone application approach. Med Nov Technol Devices 2023; 18:100237.

https://doi.org/10.1016/j.medntd.2023.100237

9. Park SM, Visbal-Onufrak MA, Haque MM, Were MC, Naanyu V, Hasan MK, et al. mHealth spectros-

copy of blood hemoglobin with spectral super-resolution. Optica 2020; 7:563. https://doi.org/10.1364/

optica.390409 PMID: 33365364

10. Dimauro G, Griseta ME, Camporeale MG, Clemente F, Guarini A, Maglietta R. An intelligent non-inva-

sive system for automated diagnosis of anemia exploiting a novel dataset. Artif Intell Med 2023;

136:102477. https://doi.org/10.1016/j.artmed.2022.102477 PMID: 36710064

11. Perez-Plazola MS, Tyburski EA, Smart LR, Howard TA, Pfeiffer A, Ware RE, et al. AnemoCheck-LRS:

an optimized, color-based point-of-care test to identify severe anemia in limited-resource settings. BMC

Med 2020; 18:337. https://doi.org/10.1186/s12916-020-01793-6 PMID: 33190639

12. Haggenmu¨ller V, Bogler L, Weber A-C, Kumar A, Ba

rnighausen T, Danquah I, et al. Smartphone-based

point-of-care anemia screening in rural Bihar in India. Commun Med 2023; 3:1–10. https://doi.org/10.

1038/s43856-023-00267-z PMID: 36949164

13. Bauskar S, Jain P, Gyanchandani M. A noninvasive computerized technique to detect anemia using

images of eye conjunctiva. Pattern Recognit Image Anal 2019; 29:438–46. https://doi.org/10.1134/

S1054661819030027

14. Chen Y-M, Miaou S-G, Bian H. Examining palpebral conjunctiva for anemia assessment with image

processing methods. Comput Methods Programs Biomed 2016; 137:125–35. https://doi.org/10.1016/j.

cmpb.2016.08.025 PMID: 28110719

15. Rivero-Palacio M, Alfonso-Morales W, Caicedo-Bravo E. Mobile application for anemia detection

through ocular conjunctiva images. 2021 IEEE Colomb. Conf. Appl. Comput. Intell. ColCACI, Cali,

Colombia: IEEE; 2021, p. 1–6. https://doi.org/10.1109/ColCACI52978.2021.9469593.

16. Mannino RG, Myers DR, Tyburski EA, Caruso C, Boudreaux J, Leong T, et al. Smartphone app for non-

invasive detection of anemia using only patient-sourced photos. Nat Commun 2018; 9:4924. https://doi.

org/10.1038/s41467-018-07262-2 PMID: 30514831

17. Korevaar DA, Cohen JF, Reitsma JB, Bruns DE, Gatsonis CA, Glasziou PP, et al. Updating standards

for reporting diagnostic accuracy: the development of STARD 2015. Res Integr Peer Rev 2016; 1:7.

https://doi.org/10.1186/s41073-016-0014-7 PMID: 29451535

18. Badireddy M, Baradhi KM. Chronic Anemia. StatPearls, Treasure Island (FL): StatPearls Publishing;

2023.

19. Transfusion thresholds adult—UpToDate n.d. https://www.uptodate.com/contents/image?imageKey=

HEME%2F93934 (accessed June 29, 2023).

20. Moretti K, Marque

s CG, Garbern S, Mbanjumucyo G, Uwamahoro C, Beaudoin FL, et al. Transfusion,

mortality and hemoglobin level: Associations among emergency department patients in Kigali, Rwanda.

Afr J Emerg Med 2020; 10:68–73. https://doi.org/10.1016/j.afjem.2020.01.004 PMID: 32612911

21. Stevens GA, Paciorek CJ, Flores-Urrutia MC, Borghi E, Namaste S, Wirth JP, et al. National, regional,

and global estimates of anaemia by severity in women and children for 2000–19: a pooled analysis of

population-representative data. Lancet Glob Health 2022; 10:e627–39. https://doi.org/10.1016/S2214-

109X(22)00084-5 PMID: 35427520

22. Anemia n.d. https://www.hematology.org/education/patients/anemia (accessed July 3, 2023).

23. America Counts Staff. RHODE ISLAND: 2020 Census. U S Census Bur 2021.

24. Safiri S, Kolahi A-A, Noori M, Nejadghaderi SA, Karamzad N, Bragazzi NL, et al. Burden of anemia and

its underlying causes in 204 countries and territories, 1990–2019: results from the Global Burden of Dis-

ease Study 2019. J Hematol OncolJ Hematol Oncol 2021; 14:185. https://doi.org/10.1186/s13045-021-

01202-2 PMID: 34736513

25. Edwards P, Zhang C, Zhang B, Hong X, Nagarajan VK, Yu B, et al. Smartphone based optical spec-

trometer for diffusive reflectance spectroscopic measurement of hemoglobin. Sci Rep 2017; 7:12224.

https://doi.org/10.1038/s41598-017-12482-5 PMID: 28939898

26. Wemyss TA, Nixon-Hill M, Outlaw F, Karsa A, Meek J, Enweronu-Laryea C, et al. Feasibility of smart-

phone colorimetry of the face as an anaemia screening tool for infants and young children in Ghana.

PloS One 2023; 18:e0281736. https://doi.org/10.1371/journal.pone.0281736 PMID: 36867642

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 12 / 13

27. Jain P, Bauskar S, Gyanchandani M. Neural network based non-invasive method to detect anemia from

images of eye conjunctiva. Int J Imaging Syst Technol 2020; 30:112–25. https://doi.org/10.1002/ima.

22359

28. Hasan MK, Haque MM, Adib R, Tumpa JF, Begum A, Love RR, et al. SmartHeLP: Smartphone-based

hemoglobin level prediction using an artificial neural network. AMIA Annu Symp Proc AMIA Symp 2018;

2018:535–44. PMID: 30815094

29. Sumner, RobertC. Processing RAW Images in MATLAB 2014. https://rcsumner.net/raw_guide/

RAWguide.pdf (accessed January 19, 2024).

30. Maslove DM, Podchiyska T, Lowe HJ. Discretization of continuous features in clinical datasets. J Am

Med Inform Assoc JAMIA 2013; 20:544–53. https://doi.org/10.1136/amiajnl-2012-000929 PMID:

23059731

31. Appiahene P, Asare JW, Donkoh ET, Dimauro G, Maglietta R. Detection of iron deficiency anemia by

medical images: a comparative study of machine learning algorithms. BioData Min 2023; 16:2. https://

doi.org/10.1186/s13040-023-00319-z PMID: 36694237

32. Young MF, Raines K, Jameel F, Sidi M, Oliveira-Streiff S, Nwajei P, et al. Non-invasive hemoglobin

measurement devices require refinement to match diagnostic performance with their high level of

usability and acceptability. PloS One 2021; 16:e0254629. https://doi.org/10.1371/journal.pone.

0254629 PMID: 34270591

33. Global digital health market size between 2015 and 2025, by major segment. Statista 2018. https://

www.statista.com/statistics/387867/value-of-worldwide-digital-health-market-forecast-by-segment/

(accessed January 23, 2024).

34. Plante TB, Urrea B, MacFarlane ZT, Blumenthal RS, Miller ER, Appel LJ, et al. Validation of the instant

blood pressure smartphone app. JAMA Intern Med 2016; 176:700. https://doi.org/10.1001/

jamainternmed.2016.0157 PMID: 26938174

35. Akbar S, Coiera E, Magrabi F. Safety concerns with consumer-facing mobile health applications and

their consequences: a scoping review. J Am Med Inform Assoc 2020; 27:330–40. https://doi.org/10.

1093/jamia/ocz175 PMID: 31599936

36. Carson JL, Stanworth SJ, Roubinian N, Fergusson DA, Triulzi D, Doree C, et al. Transfusion thresholds

and other strategies for guiding allogeneic red blood cell transfusion. Cochrane Database Syst Rev

2016; 10:CD002042. https://doi.org/10.1002/14651858.CD002042.pub4 PMID: 27731885

37. Addo OY, Yu EX, Williams AM, Young MF, Sharma AJ, Mei Z, et al. Evaluation of hemoglobin cutoff lev-

els to define anemia among healthy individuals. JAMA Netw Open 2021; 4:e2119123–e2119123.

https://doi.org/10.1001/jamanetworkopen.2021.19123 PMID: 34357395

38. Cappellini MD, Motta I. Anemia in Clinical Practice-Definition and Classification: Does hemoglobin

change with aging? Semin Hematol 2015; 52:261–9. https://doi.org/10.1053/j.seminhematol.2015.07.

006 PMID: 26404438

39. Skelton VA, Wijayasinghe N, Sharafudeen S, Sange A, Parry NS, Junghans C. Evaluation of point-of-

care haemoglobin measuring devices: a comparison of Radical-7

pulse co-oximetry, HemoCue

and laboratory haemoglobin measurements in obstetric patients*: Evaluation of point-of-care haemo-

globin measuring devices. Anaesthesia 2013; 68:40–5. https://doi.org/10.1111/anae.12039 PMID:

23088815

40. Hornedo-Gonza

lez KD, Jacob AK, Burt JM, Higgins AA, Engel EM, Hanson AC, et al. Non-invasive

hemoglobin estimation for preoperative anemia screening. Transfusion (Paris) 2023; 63:315–22.

https://doi.org/10.1111/trf.17237 PMID: 36605019

PLOS ONE

Anemia prediction using a smartphone camera

PLOS ONE | https://doi.org/10.1371/journal.pone.0302883 May 13, 2024 13 / 13