Smarter Fortification: Leveraging AI to Detect Iron Levels in Wheat Flour

Iron deficiency doesn’t make headlines, but it should. Silently impacting millions across India, especially women and young children, iron deficiency is one of the main causes of anaemia and a major contributor to maternal deaths. The numbers speak for themselves: nearly three in five children under five, and half of all women of reproductive age, are anaemic. 

One of the most cost-effective ways to tackle this problem is by adding iron to everyday foods like wheat flour - a strategy backed by the World Health Organization. But there’s a catch: making sure the right amount of iron is added to the flour is no easy task. While there are advanced lab tests that can measure iron accurately, they’re expensive, time-consuming, and require skilled personnel. On the other hand, cheaper tests such as the iron spot test can be easily done at mills but rely heavily on human judgment, which can lead to errors. 

To bridge this gap, Fortify Health turned to technology to explore whether AI could help. In partnership with a graduate student from  Georgia Institute of Technology (HaLim Jun), the team developed an early prototype of an AI-based solution in 2023. It uses simple images from the iron spot test and machine learning to estimate iron content more reliably. Encouraged by the advances in AI, Fortify Health collaborated with Hornbill Agritech, an AI specialist in food and agriculture, to develop a robust AI product that enhances quality control, making it smarter, faster, and more scalable, a crucial step toward reducing iron deficiency across the country. 

This is the first in Fortify Health’s series of blogs on our experience with AI. 

Enhancing Iron Fortification: AI for Accurate and Scalable Quality Control

For many mothers in India, the consequences of iron deficiency and anaemia are life-threatening. Anaemia is a major contributor to maternal mortality in India, where the rate stands at 97 deaths per 100,000 live births. In 2020 alone, this resulted in around 20,000 women losing their lives due to pregnancy-related causes. 
​
Recognising the scale of the problem, the Government of India launched an initiative called Anaemia Mukt Bharat (Anaemia-Free India), which promotes a range of strategies to combat anaemia, including fortification of staple foods like wheat flour with iron and other essential micronutrients.

In India, wheat flour (atta) is the second most widely consumed staple food, eaten daily by millions. It is one of the most powerful vehicles for delivering iron at scale. However, the effectiveness of flour fortification depends heavily on one crucial factor: ensuring the correct amount of iron is added in line with the  Food Safety and Standards Authority of India (FSSAI)’s regulatory standards. If iron levels are inconsistent or incorrect, fortified foods may lose their intended health benefits and, in cases of excess, even pose safety risks. 

Despite this, flour producers often rely on a manual ‘iron spot test,’ which is quick and low-cost but can be inaccurate since it depends on human judgment. More accurate lab tests exist, but are expensive and slow, often taking over a week to deliver results.

To advance quality control for iron-fortified flour, we have developed a machine-learning-based image processing pipeline that improves accuracy in measuring iron content in fortified wheat flour.. Our AI solution analyses images from the iron spot test, a quick but low-accuracy method, to predict iron levels with greater accuracy and objectivity. It can flag samples that fall outside safe ranges early, making quality control not only more precise but also faster and more affordable, while removing delays caused by laboratory testing.

How Our AI Measures Iron in Flour or AI Methodology: Image Processing and Iron Content Estimation

Image processing pipeline

An iron spot test image is first cropped and brightness-normalised to eliminate unnecessary background images and lighting bias; regions above a calibrated brightness threshold are masked to suppress glare. We then apply OpenCV’s multi-scale blob detector to isolate the iron-reaction spots and extract key features such as count and size. These descriptors feed a supervised learning model that estimates the flour’s iron concentration in real time. 

We tested two different models: 

• ⦿ A regression model that predicts the numerical concentration of iron in the flour.

• ⦿ A ​classification model that determines whether the iron content falls within the regulatory standards (14-21.25 mg/kg)

We tested five different models, including Gradient Boosting Regressor, K-neighbors regressor, Randomforest, Catboost Regressor, LGBM regressor, and support vector machine and statistical model for confidence interval.

We utilised a total of 213 datasets consisting of iron spot test images paired with their corresponding ground truth iron content. These datasets were created using two distinct methodologies:

• ⦿ Operational sampling, where real-world samples were tested using the quantitative iCheck method; and

• ⦿ Synthetic data generation, where controlled iron-flour mixtures were prepared and their corresponding iron spot tests were conducted. The synthetic data generation process allowed for efficient dataset creation by bypassing operational sampling. It also eliminated the need for the quantitative iCheck test, making the overall data collection process more time- and cost-effective.

Improved Accuracy and Impact: AI-Enhanced Quality Control 

High accuracy was achieved when training solely on the operational samples, with an error rate of just 12%. However, as more synthetic data was introduced, model performance declined. This was due to the synthetic samples including a wider range of iron concentrations, many of which rarely appear in real-world operational settings. This broader range caused a weaker correlation between the image features and the actual iron content.

To address this, we adopted a multi-model classification strategy by overlaying two binary classifiers: one specialised in detecting whether a sample exceeds the safe iron concentration threshold, and another that identifies whether it falls below. By combining multiple models, we made the system more reliable, and it correctly identified results nearly 8 out of 10 times.[1]. (F1score - 77.8%) 

This innovative approach directly supports global and national guidelines recommending flour fortification as a cost-effective strategy to enhance nutrition and combat micronutrient deficiencies. By improving the reliability and efficiency of quality assurance, Fortify Health strengthens its ability to ensure regulatory compliance with standards like those set by the Food Safety and Standards Authority of India (FSSAI). This has a direct and meaningful impact on the lives of millions of people across India, particularly in regions where Fortify Health is actively working with 160+ open market millers to produce more than 40,000 MT of wheat flour each month. 

Next Steps: Scaling AI for Reliable Fortification 

Given the success of the early pilot, Fortify Health is now collaborating with Hornbill Agritech to build on this momentum and develop a large-scale, validated version of the AI tool. The goal is to develop a reliable and scalable product that can be made available as a public good to enhance quality control in wheat flour fortification. 

While this technology represents a significant advancement, like any new system, there are areas for continued refinement. Limitations identified include the need for ongoing training with new data, especially for flour samples with higher iron content. This is necessary to ensure accurate prediction across the full regulatory range. Furthermore, environmental variations during image capture, such as differing lighting, can affect the model's performance and require ongoing attention and mitigation strategies. Addressing these points through continued development and collaboration will be key to maximising the tool's potential.

To learn more about updates on this pioneering model, its potential as a digital public good, and how Fortify Health is leveraging technology to improve public health outcomes in India, visit Fortify Health’s website. You can also subscribe to our newsletter to stay updated on the progress of this project and to learn about Fortify Health's broader efforts to combat micronutrient deficiencies.

---

Sources:

-  India - One of the countries with the highest anaemia, with over 55% of children (children between 6-59 months)  having anaemia. (Source)
- Anaemia remains a critical public health issue in India. According to the National Family Health Survey (NFHS-5, 2019–2021), approximately 67% of children under five and 57% of women of reproductive age are anaemic.
-A major implication related to iron deficiency is maternal death, which was  97 per 100,000 live births in 2018–20. (Source)
-It means that with an estimated 20 thousand maternal deaths in 2020 (Link)

---