HemoglobinAI

Development of a Smartphone-Based Machine Learning Pipeline for Hemoglobin Level Prediction among Pregnant Women: A Comparative Performance of Palpebral Conjunctiva with Buccal Mucosa Images

Dr. Jannatbi L. Iti¹, Anirudh Gangadharan²
¹Assistant Professor, Department of Community Medicine · ²IIIrd Year Part 1 MBBS Student
Gadag Institute of Medical Sciences, Karnataka
AFMC Oral Presentation · May 1, 2026

Background

Anemia affects 52.2% of pregnant women in India.¹
Anaemia is directly or indirectly responsible for 40% of maternal mortality, despite national programs such as Anemia Mukt Bharat and Poshan Abhiyan.²
Routine haemoglobin estimation relies on invasive methods that impose compounded costs: transport, laboratory fees, and lost wages.³
Conjunctival and mucosal pallor are established clinical indicators of anemia, representing a non-invasive signal amenable to smartphone imaging.⁴

Objectives

To develop a machine learning model for haemoglobin level prediction using smartphone images from palpebral conjunctiva and buccal mucosa among pregnant women at a tertiary care centre.
To compare machine learning models based on palpebral conjunctiva images with those based on buccal mucosa images for haemoglobin level prediction among pregnant women at a tertiary care centre.

Methodology

Study design. Observational cross-sectional study. Source population: pregnant women aged 18–45 attending OBG OPD for routine antenatal checkup at GIMS, Gadag. Study period: 1 January 2025 to 31 December 2025.

Sample size. n = z²pq/d² = 600, computed using prevalence p = 46%,⁵ q = 54, d = 9% of p = 4.14. The same n = 600 satisfies 5-fold cross-validation requirements. Sampling technique: purposive sampling.

Inclusion. All pregnant women attending the centre who gave written informed consent.

Exclusion. Active eye infections (e.g. conjunctivitis); cosmetic contact lenses; periorbital trauma; blood transfusion or IV iron sucrose within the preceding 15 days.

Data collection. Following IEC approval and written informed consent, data were collected via pre-tested semi-structured questionnaire, including standardised smartphone photographs of the palpebral conjunctiva and buccal mucosa alongside a colour-calibration reference card.

Statistical analysis. Data entered in MS Excel; analysed using frequency, percentage, and chi-square test (SPSS v21).

End-to-end ML pipeline: smartphone image capture with colour calibration card → YOLOv8 segmentation → feature extraction (RGB, HSV, LAB + sociodemographic) → ML model ensemble → performance metrics — Figure 1. End-to-end pipeline. Smartphone images of the palpebral conjunctiva and buccal mucosa are colour-calibrated using a reference card, then segmented with a fine-tuned YOLOv8 model. RGB, HSV, and LAB colour features are combined with 52 sociodemographic and clinical features before evaluation across seven ML algorithms with 5-fold cross-validation.

Results

Among 600 pregnant women, mean age was 25.34 ± 3.889 years. The age distribution and haemoglobin summary statistics are shown below.

Bar chart: distribution by age group (n=600) and Table 01 showing haemoglobin means and standard deviations — Figure 2 & Table 01. Left: distribution by age group — 356 participants (59.3%) were ≤25 years, 238 (39.7%) were 26–35 years, and 6 (1.0%) were >35 years. Right: mean haemoglobin by measurement method.

The chi-square analysis of the association between laboratory haemoglobin and predicted haemoglobin by imaging site is shown in Table 02.

Table 02. Association between cyanmethemoglobin Hb and predicted Hb levels by imaging site (n = 600)
Hb by cyanmethemoglobin	Palpebral Conjunctiva				Buccal Mucosa
	<10 g/dL	10–11 g/dL	>11 g/dL	Total	<10 g/dL	10–11 g/dL	>11 g/dL	Total
<10 g/dL	0.4% (2)	0.0% (0)	0.0% (0)	0.4% (2)	0.4% (2)	0.0% (0)	0.2% (1)	0.5% (3)
10–11 g/dL	9.8% (59)	15.3% (92)	20.1% (124)	45.8% (275)	9.4% (57)	14.0% (84)	20.3% (122)	43.8% (263)
>11 g/dL	9.8% (59)	16.7% (100)	27.3% (164)	53.8% (323)	10.2% (61)	18.0% (108)	27.5% (165)	55.7% (334)
Total	20.0% (120)	32.0% (192)	48.0% (288)	100% (600)	20.0% (120)	32.0% (192)	48.0% (288)	100% (600)
p value	0.04 (χ² = 10.1, df = 4)				0.238 (χ² = 5.524, df = 4)

Palpebral conjunctiva shows a statistically significant association with laboratory Hb (p = 0.04); buccal mucosa does not (p = 0.238).

Table 03 reports model performance across both imaging sites. SVR achieves the highest ROC AUC on both sites; palpebral conjunctiva consistently outperforms buccal mucosa across all models and all metrics.

Table 03. ML model performance — palpebral conjunctiva vs buccal mucosa (n = 600)
Model	Palpebral Conjunctiva				Buccal Mucosa
Model	AUC	F1	Acc.	Prec.	AUC	F1	Acc.	Prec.
Ridge	0.617	0.514	0.583	0.473	0.602	0.520	0.585	0.475
Lasso	0.505	0.179	0.573	0.364	0.497	0.211	0.575	0.386
ElasticNet	0.549	0.392	0.565	0.433	0.547	0.413	0.565	0.438
SVR ★	0.636	0.522	0.593	0.484	0.609	0.519	0.598	0.489
Random Forest	0.610	0.515	0.588	0.478	0.593	0.507	0.592	0.481
Gradient Boosting	0.599	0.508	0.583	0.473	0.571	0.491	0.575	0.462
XGBoost	0.604	0.518	0.585	0.475	0.590	0.496	0.570	0.458

★ Best model by ROC AUC. Acc. = Accuracy, Prec. = Precision.

ROC curves for anemia detection (Hb < 11 g/dL) — Panel A: Palpebral Conjunctiva, Panel B: Buccal Mucosa, seven models each — Figure 3. ROC curves for anemia detection (<11 g/dL). A: Palpebral conjunctiva — SVR achieves highest AUC = 0.636. B: Buccal mucosa — SVR AUC = 0.609. Palpebral conjunctiva shows consistently superior discrimination across all seven models.

Conclusion

Machine learning models based on palpebral conjunctiva and buccal mucosa smartphone images can predict haemoglobin levels in pregnant women.
Palpebral conjunctiva performs marginally but consistently better than buccal mucosa across all models and metrics.
SVR outperforms the other six algorithms (Ridge, Lasso, ElasticNet, Random Forest, Gradient Boosting, XGBoost) on both imaging sites.
Smartphone-based non-invasive screening using either imaging site has potential as an anemia screening tool in low-resource settings.

Recommendations

A multicentric design and a longitudinal follow-up study are the natural next steps to validate generalisability across populations and smartphone hardware.

Live Demo

Upload a conjunctival image, enter clinical parameters, and receive an instant haemoglobin estimate. Open on HuggingFace →

References

International Institute for Population Sciences (IIPS) and ICF. National Family Health Survey (NFHS-5), 2019–21: India. Mumbai: IIPS; 2021.
Singh A, Mangal M. Laboratory methods for estimation of hemoglobin: A review. Int J Adv Med. 2018;5(4):877–881.
Mannino RG, Myers DR, Ahn B, et al. Smartphone app for non-invasive detection of anemia using only patient-sourced photos. Nat Commun. 2018;9(1):4924.
Suner S, Parthasarathy S, Kasikcioglu E. Prediction of anemia from photographs of the palpebral conjunctiva. JAMA Netw Open. 2021;4(5):e215599.
International Institute for Population Sciences (IIPS) and ICF. National Family Health Survey (NFHS-4), India, 2015–16: Karnataka. Mumbai: IIPS; 2017.
Jocher G, Chaurasia A, Qiu J. Ultralytics YOLO [v8.0.0]. 2023. github.com/ultralytics/ultralytics

The authors thank the Director, Principal, and Head of Department for permission to conduct this study; PG residents and interns for assistance with data collection; and all participants for their willing participation.