Clinical Study · Maternal Health · AI

HemoglobinAI

Smartphone-Based Machine Learning Pipeline for Non-Invasive Hemoglobin Estimation — Comparative Performance of Palpebral Conjunctiva and Buccal Mucosa in 600 Pregnant Women
Anirudh Gangadharan & collaborators
KHPIMS, Gadag, Karnataka · Cross-sectional study · n = 600
600
Pregnant Women
1.018
Best MAE (g/dL)
0.610
AUC-ROC
7
ML Models Tested

Anaemia in pregnancy remains a critical public health challenge in India, affecting nearly 50% of pregnant women and contributing directly to maternal and neonatal mortality. The gold standard — laboratory haemoglobin measurement — requires venipuncture, trained phlebotomists, and laboratory infrastructure, making frequent monitoring infeasible in resource-limited settings.

In rural primary health centres and community outreach camps, the choice is often binary: perform invasive testing (expensive, slow, requires cold chain) or rely on clinical pallor assessment (subjective, sensitivity <60%). A non-invasive, smartphone-based screening tool could bridge this gap, enabling community health workers to triage pregnant women for anaemia without drawing blood.

1

Standardised Image Capture

Smartphone photographs of the palpebral conjunctiva (required) and buccal mucosa (optional) are captured alongside a colour calibration card with known colour patches, enabling white-balance normalisation across different smartphones and lighting conditions.

2

YOLOv8 Region-of-Interest Segmentation

A fine-tuned YOLOv8 model automatically segments conjunctival or mucosal tissue from surrounding anatomy, isolating the diagnostically relevant region. Calibration card patches are simultaneously detected for colour normalisation.

3

Multi-Modal Feature Extraction

Image features (mean RGB, colour histograms, texture metrics) are combined with 52 sociodemographic and clinical features: age, dietary patterns, obstetric history, and medical history — yielding a 61-feature vector per patient (70 when both sites are combined).

4

ML Model Ensemble & Prediction

Seven algorithms (Ridge, Lasso, ElasticNet, SVR, Random Forest, Gradient Boosting, XGBoost) are evaluated with 5-fold cross-validation. The model outputs a continuous haemoglobin estimate (g/dL) with severity classification.

Conjunctival image capture with colour calibration card
Conjunctival capture + demographics
Buccal mucosa capture and medical history
Buccal mucosa + medical history
Obstetric history input
Obstetric history input
Estimated hemoglobin result
Result: Hb 11.4 g/dL — Mild Anaemia

Try the Live Demo

Upload a conjunctival image, enter clinical parameters, and receive an instant haemoglobin estimate.

Open on HuggingFace →

Among 600 pregnant women (mean age 25.34 ± 3.89 years), 32.17% were primigravida and 39.2% were anaemic by laboratory confirmation. Palpebral conjunctiva imaging with Random Forest regression achieved the best overall performance.

Best model: Random Forest on palpebral conjunctiva — MAE = 1.018 g/dL, AUC-ROC = 0.610, with 57.3% of predictions within ±1.0 g/dL of laboratory values.

Clinical Accuracy — Palpebral Conjunctiva

32.5%
± 0.5 g/dL
57.3%
± 1.0 g/dL
76.2%
± 1.5 g/dL
87.8%
± 2.0 g/dL

Site Comparison (Random Forest)

Imaging Site MAE (g/dL) AUC-ROC Sensitivity Specificity
Palpebral Conjunctiva 1.018 0.022 0.610 0.557 0.608
Buccal Mucosa 1.039 0.014 0.593 0.536 0.627
Combined 1.028 0.017 0.599 0.549 0.600
Performance comparison across imaging sites and models
Figure 1. Performance comparison across imaging sites and ML models. Palpebral conjunctiva shows marginally superior AUC-ROC and lower MAE. SVR achieved highest AUC-ROC (0.636) while Random Forest achieved lowest MAE (1.018 g/dL).

Honest assessment: The low R² values (0.014–0.022) indicate that the current pipeline explains very little variance in haemoglobin levels. The primary contributions are the validated 600-patient paired dataset, the end-to-end pipeline infrastructure, and the comparative evidence that conjunctival imaging outperforms buccal mucosa. These represent a foundation for deeper architectures, not a deployment-ready diagnostic.

Despite modest model performance, the contribution is infrastructural. This work establishes a reproducible, colour-calibrated capture protocol replicable across any smartphone; a curated dataset of 600 paired image–laboratory observations from a real clinical cohort; and evidence that palpebral conjunctiva provides a stronger imaging signal than buccal mucosa, guiding future research toward the more informative anatomical site.

In the context of postpartum haemorrhage triage, where any haemoglobin estimate — even approximate — outperforms having no information at all, a screening tool with MAE of ~1 g/dL could differentiate severe anaemia from normal levels, enabling triage decisions that currently rely on subjective pallor assessment alone.

Dataset & Infrastructure Contribution

600 standardised, colour-calibrated smartphone images paired with laboratory haemoglobin values from pregnant women in a resource-limited Indian clinical setting — a dataset designed for foundation model fine-tuning and transfer learning that could substantially improve on the classical ML baseline reported here.

The immediate next step is applying pretrained medical vision foundation models (BiomedCLIP, PubMedCLIP) via transfer learning, replacing handcrafted colour features with deep learned representations. We are also exploring whether a vision transformer fine-tuned on conjunctival images alone — without sociodemographic features — can outperform the current multi-modal pipeline, simplifying deployment to a single-photo screening tool.

Longer-term goals include multi-site validation across diverse populations and smartphone hardware, integration with the COGNIT semantic protocol for real-time triage transmission, and a prospective clinical trial comparing smartphone-based screening against standard laboratory testing in community health worker settings.