Proof of Good | SDG2: Zero Hunger

Leveraging Machine Learning to Forecast Child Malnutrition (Case Study: Kenya)

In the global effort to achieve the United Nations Sustainable Development Goals (SDGs), predictive data systems are emerging as powerful tools to prevent humanitarian crises before they escalate. One compelling example is Kenya’s use of artificial intelligence (AI) to forecast child malnutrition rates up to six months in advance, directly contributing to SDG 2: Zero Hunger.

This initiative demonstrates how machine learning, when combined with health system data and satellite environmental indicators, can shift governments from reactive emergency response to proactive prevention, enabling earlier interventions that protect vulnerable children before malnutrition reaches critical levels.

Context and Challenge

Child malnutrition remains a serious and persistent challenge resulting from the hunger crisis in Kenya, particularly in arid and semi-arid regions. Acute malnutrition occurs when children lose weight rapidly due to insufficient food intake or illness. It significantly increases the risk of disease and death among children under five.

In Kenya, malnutrition rates fluctuate due to multiple reasons, including: drought cycles, food price volatility, disease outbreaks, and climate variability.

Historically, the Ministry of Health relied on routine health facility reporting through Kenya’s District Health Information System (DHIS2), which aggregates data from more than 17,000 health facilities nationwide. While this system provides extensive monthly data on malnutrition cases, diarrhoeal disease, and low birth weight, traditional forecasting methods based on historical averages have limitations.

 

These methods assume past trends will continue, making it difficult to anticipate sudden surges triggered by environmental shocks or food insecurity. As a result, interventions often begin after malnutrition rates have already reached critical levels, reducing their effectiveness and increasing the risk of child mortality.

Kenya needed a system capable of forecasting risk early enough to allow preventative action.

Data-Driven Solution

To address this gap, researchers from the University of Southern California (USC) partnered with the Kenya Ministry of Health, Amref Health Africa, and the Microsoft AI

for Good Lab to develop a machine learning model capable of forecasting child malnutrition at the sub-county level.

The model integrates two primary categories of data:

  1. Health System Data such as monthly reports from over 17,000 health facilities, recorded cases of acute malnutrition, and indicators such as diarrhoeal disease treatment and low birth weight.
  2. Satellite Environmental Data such as measurements of gross primary productivity (GPP), which is an indicator of vegetation health and crop growth and environmental signals reflecting drought and agricultural stress.

By combining human health indicators with environmental data, the model captures both immediate clinical trends and broader food system pressures.

Machine Learning Methodology:

Researchers tested multiple modeling approaches and found that gradient boosting machine learning models delivered the strongest predictive performance.

The model was trained to:

detect complex relationships between environmental shocks and malnutrition rates, forecast acute malnutrition prevalence at the sub-county level, and predict risk up to six months in advance.

This multi-variable, data-integrated approach significantly outperformed traditional forecasting techniques.

Implementation and Results

The machine learning model demonstrated strong predictive accuracy:

  • Approximately 89% accuracy for one-month-ahead forecasts
  • Approximately 86% accuracy for six-month-ahead forecasts

Compared to Kenya’s standard “window average” forecasting method, the machine learning model showed substantially improved performance, particularly in regions where malnutrition rates fluctuate unpredictably. The model’s predictive strength was measured using area-under-the-curve (AUC) metrics and it scored 0.86 AUC over six months compared to 0.73 AUC of the traditional baseline model.

This improvement represents a significant advancement in the ability to identify high-risk areas before malnutrition becomes widespread.

To translate these forecasts into actionable policy tools, researchers developed a decision-support dashboard that visualizes predicted malnutrition risk at the sub-county level. The dashboard allows the Kenya Ministry of Health and partner organizations to identify emerging hotspots, allocate therapeutic foods and supplements earlier, deploy community health workers strategically, and eventually adjust nutrition programming proactively.

Outcomes and Impact

  • Early warning capability and improved resource allocation:
    The Kenya case represents a fundamental shift from reactive crisis management to predictive prevention. By forecasting malnutrition risk up to six months in advance, authorities gain critical lead time to mobilize resources before children reach severe wasting stages. While, high-accuracy geographic targeting enables more efficient distribution of ready-to-use therapeutic foods (RUTF), nutrition supplements, and community health interventions.
  • Health system strengthening and scalability
    The initiative leverages Kenya’s existing DHIS2 infrastructure, demonstrating how national health data systems can be augmented with AI rather than replaced.

    DHIS2 is used in more than 80 low- and middle-income countries, meaning this forecasting framework has strong potential for replication globally.

    Rather than relying solely on emergency appeals after malnutrition rates spike, this approach provides a data-informed mechanism for anticipatory action aiming to mitigate the impact of the hunger crisis.

Conclusion

Kenya’s AI-powered malnutrition forecasting system exemplifies how machine learning can directly support SDG 2: Zero Hunger. By integrating health facility data from over 17,000 reporting centers with satellite-derived environmental indicators, researchers built a predictive model capable of forecasting acute child malnutrition with up to 89% short-term accuracy and 86% six-month accuracy.

This initiative shows that data is not merely descriptive but it can also be predictive, preventative, and life-saving. Through collaboration between academic researchers, government institutions, global health organizations, and AI specialists, Kenya has pioneered a scalable model for anticipatory nutrition response, proving that the ability to act before crisis peaks is as important as the ability to respond after it begins.

Leveraging Machine Learning to Forecast Child Malnutrition (Case Study: Kenya)

In the global effort to achieve the United Nations Sustainable Development Goals (SDGs), predictive data systems are emerging as powerful tools to prevent humanitarian crises before they escalate. One compelling example is Kenya’s use of artificial intelligence (AI) to forecast child malnutrition rates up to six months in advance, directly contributing to SDG 2: Zero Hunger.

This initiative demonstrates how machine learning, when combined with health system data and satellite environmental indicators, can shift governments from reactive emergency response to proactive prevention, enabling earlier interventions that protect vulnerable children before malnutrition reaches critical levels.

Context and Challenge

Child malnutrition remains a serious and persistent challenge resulting from the hunger crisis in Kenya, particularly in arid and semi-arid regions. Acute malnutrition occurs when children lose weight rapidly due to insufficient food intake or illness. It significantly increases the risk of disease and death among children under five.

In Kenya, malnutrition rates fluctuate due to multiple reasons, including: drought cycles, food price volatility, disease outbreaks, and climate variability.

Historically, the Ministry of Health relied on routine health facility reporting through Kenya’s District Health Information System (DHIS2), which aggregates data from more than 17,000 health facilities nationwide. While this system provides extensive monthly data on malnutrition cases, diarrhoeal disease, and low birth weight, traditional forecasting methods based on historical averages have limitations.

These methods assume past trends will continue, making it difficult to anticipate sudden surges triggered by environmental shocks or food insecurity. As a result, interventions often begin after malnutrition rates have already reached critical levels, reducing their effectiveness and increasing the risk of child mortality.

Kenya needed a system capable of forecasting risk early enough to allow preventative action.

Data-Driven Solution

To address this gap, researchers from the University of Southern California (USC) partnered with the Kenya Ministry of Health, Amref Health Africa, and the Microsoft AI

for Good Lab to develop a machine learning model capable of forecasting child malnutrition at the sub-county level.

The model integrates two primary categories of data:

    1. Health System Data such as monthly reports from over 17,000 health facilities, recorded cases of acute malnutrition, and indicators such as diarrhoeal disease treatment and low birth weight.
    2. Satellite Environmental Data such as measurements of gross primary productivity (GPP), which is an indicator of vegetation health and crop growth and environmental signals reflecting drought and agricultural stress.

By combining human health indicators with environmental data, the model captures both immediate clinical trends and broader food system pressures.

Machine Learning Methodology:

Researchers tested multiple modeling approaches and found that gradient boosting machine learning models delivered the strongest predictive performance.

The model was trained to:

detect complex relationships between environmental shocks and malnutrition rates, forecast acute malnutrition prevalence at the sub-county level, and predict risk up to six months in advance.

This multi-variable, data-integrated approach significantly outperformed traditional forecasting techniques.

Implementation and Results

The machine learning model demonstrated strong predictive accuracy:

  • Approximately 89% accuracy for one-month-ahead forecasts
  • Approximately 86% accuracy for six-month-ahead forecasts

Compared to Kenya’s standard “window average” forecasting method, the machine learning model showed substantially improved performance, particularly in regions where malnutrition rates fluctuate unpredictably. The model’s predictive strength was measured using area-under-the-curve (AUC) metrics and it scored 0.86 AUC over six months compared to 0.73 AUC of the traditional baseline model.

This improvement represents a significant advancement in the ability to identify high-risk areas before malnutrition becomes widespread.

To translate these forecasts into actionable policy tools, researchers developed a decision-support dashboard that visualizes predicted malnutrition risk at the sub-county level. The dashboard allows the Kenya Ministry of Health and partner organizations to identify emerging hotspots, allocate therapeutic foods and supplements earlier, deploy community health workers strategically, and eventually adjust nutrition programming proactively.

Outcomes and Impact

  • Early warning capability and improved resource allocation:
    The Kenya case represents a fundamental shift from reactive crisis management to predictive prevention. By forecasting malnutrition risk up to six months in advance, authorities gain critical lead time to mobilize resources before children reach severe wasting stages. While, high-accuracy geographic targeting enables more efficient distribution of ready-to-use therapeutic foods (RUTF), nutrition supplements, and community health interventions.
  • Health system strengthening and scalability:
    The initiative leverages Kenya’s existing DHIS2 infrastructure, demonstrating how national health data systems can be augmented with AI rather than replaced.

    DHIS2 is used in more than 80 low- and middle-income countries, meaning this forecasting framework has strong potential for replication globally.

    Rather than relying solely on emergency appeals after malnutrition rates spike, this approach provides a data-informed mechanism for anticipatory action aiming to mitigate the impact of the hunger crisis.

Conclusion

Kenya’s AI-powered malnutrition forecasting system exemplifies how machine learning can directly support SDG 2: Zero Hunger. By integrating health facility data from over 17,000 reporting centers with satellite-derived environmental indicators, researchers built a predictive model capable of forecasting acute child malnutrition with up to 89% short-term accuracy and 86% six-month accuracy.

This initiative shows that data is not merely descriptive but it can also be predictive, preventative, and life-saving. Through collaboration between academic researchers, government institutions, global health organizations, and AI specialists, Kenya has pioneered a scalable model for anticipatory nutrition response, proving that the ability to act before crisis peaks is as important as the ability to respond after it begins.

Leveraging Machine Learning to Forecast Child Malnutrition (Case Study: Kenya)

In the global effort to achieve the United Nations Sustainable Development Goals (SDGs), predictive data systems are emerging as powerful tools to prevent humanitarian crises before they escalate. One compelling example is Kenya’s use of artificial intelligence (AI) to forecast child malnutrition rates up to six months in advance, directly contributing to SDG 2: Zero Hunger.

This initiative demonstrates how machine learning, when combined with health system data and satellite environmental indicators, can shift governments from reactive emergency response to proactive prevention, enabling earlier interventions that protect vulnerable children before malnutrition reaches critical levels.

Context and Challenge

Child malnutrition remains a serious and persistent challenge resulting from the hunger crisis in Kenya, particularly in arid and semi-arid regions. Acute malnutrition occurs when children lose weight rapidly due to insufficient food intake or illness. It significantly increases the risk of disease and death among children under five.

In Kenya, malnutrition rates fluctuate due to multiple reasons, including: drought cycles, food price volatility, disease outbreaks, and climate variability.

Historically, the Ministry of Health relied on routine health facility reporting through Kenya’s District Health Information System (DHIS2), which aggregates data from more than 17,000 health facilities nationwide. While this system provides extensive monthly data on malnutrition cases, diarrhoeal disease, and low birth weight, traditional forecasting methods based on historical averages have limitations.

 

These methods assume past trends will continue, making it difficult to anticipate sudden surges triggered by environmental shocks or food insecurity. As a result, interventions often begin after malnutrition rates have already reached critical levels, reducing their effectiveness and increasing the risk of child mortality.

Kenya needed a system capable of forecasting risk early enough to allow preventative action.

Data-Driven Solution

To address this gap, researchers from the University of Southern California (USC) partnered with the Kenya Ministry of Health, Amref Health Africa, and the Microsoft AI

for Good Lab to develop a machine learning model capable of forecasting child malnutrition at the sub-county level.

The model integrates two primary categories of data:

  1. Health System Data such as monthly reports from over 17,000 health facilities, recorded cases of acute malnutrition, and indicators such as diarrhoeal disease treatment and low birth weight.
  2. Satellite Environmental Data such as measurements of gross primary productivity (GPP), which is an indicator of vegetation health and crop growth and environmental signals reflecting drought and agricultural stress.

By combining human health indicators with environmental data, the model captures both immediate clinical trends and broader food system pressures.

Machine Learning Methodology:

Researchers tested multiple modeling approaches and found that gradient boosting machine learning models delivered the strongest predictive performance.

The model was trained to:

detect complex relationships between environmental shocks and malnutrition rates, forecast acute malnutrition prevalence at the sub-county level, and predict risk up to six months in advance.

This multi-variable, data-integrated approach significantly outperformed traditional forecasting techniques.

Implementation and Results

The machine learning model demonstrated strong predictive accuracy:

  • Approximately 89% accuracy for one-month-ahead forecasts
  • Approximately 86% accuracy for six-month-ahead forecasts

Compared to Kenya’s standard “window average” forecasting method, the machine learning model showed substantially improved performance, particularly in regions where malnutrition rates fluctuate unpredictably. The model’s predictive strength was measured using area-under-the-curve (AUC) metrics and it scored 0.86 AUC over six months compared to 0.73 AUC of the traditional baseline model.

This improvement represents a significant advancement in the ability to identify high-risk areas before malnutrition becomes widespread.

To translate these forecasts into actionable policy tools, researchers developed a decision-support dashboard that visualizes predicted malnutrition risk at the sub-county level. The dashboard allows the Kenya Ministry of Health and partner organizations to identify emerging hotspots, allocate therapeutic foods and supplements earlier, deploy community health workers strategically, and eventually adjust nutrition programming proactively.

Outcomes and Impact

  • Early warning capability and improved resource allocation:
    The Kenya case represents a fundamental shift from reactive crisis management to predictive prevention. By forecasting malnutrition risk up to six months in advance, authorities gain critical lead time to mobilize resources before children reach severe wasting stages. While, high-accuracy geographic targeting enables more efficient distribution of ready-to-use therapeutic foods (RUTF), nutrition supplements, and community health interventions.
  • Health system strengthening and scalability
    The initiative leverages Kenya’s existing DHIS2 infrastructure, demonstrating how national health data systems can be augmented with AI rather than replaced.

    DHIS2 is used in more than 80 low- and middle-income countries, meaning this forecasting framework has strong potential for replication globally.

    Rather than relying solely on emergency appeals after malnutrition rates spike, this approach provides a data-informed mechanism for anticipatory action aiming to mitigate the impact of the hunger crisis.

Conclusion

Kenya’s AI-powered malnutrition forecasting system exemplifies how machine learning can directly support SDG 2: Zero Hunger. By integrating health facility data from over 17,000 reporting centers with satellite-derived environmental indicators, researchers built a predictive model capable of forecasting acute child malnutrition with up to 89% short-term accuracy and 86% six-month accuracy.

This initiative shows that data is not merely descriptive but it can also be predictive, preventative, and life-saving. Through collaboration between academic researchers, government institutions, global health organizations, and AI specialists, Kenya has pioneered a scalable model for anticipatory nutrition response, proving that the ability to act before crisis peaks is as important as the ability to respond after it begins.

Leveraging Machine Learning to Forecast Child Malnutrition (Case Study: Kenya)

In the global effort to achieve the United Nations Sustainable Development Goals (SDGs), predictive data systems are emerging as powerful tools to prevent humanitarian crises before they escalate. One compelling example is Kenya’s use of artificial intelligence (AI) to forecast child malnutrition rates up to six months in advance, directly contributing to SDG 2: Zero Hunger.

This initiative demonstrates how machine learning, when combined with health system data and satellite environmental indicators, can shift governments from reactive emergency response to proactive prevention, enabling earlier interventions that protect vulnerable children before malnutrition reaches critical levels.

Context and Challenge

Child malnutrition remains a serious and persistent challenge resulting from the hunger crisis in Kenya, particularly in arid and semi-arid regions. Acute malnutrition occurs when children lose weight rapidly due to insufficient food intake or illness. It significantly increases the risk of disease and death among children under five.

In Kenya, malnutrition rates fluctuate due to multiple reasons, including: drought cycles, food price volatility, disease outbreaks, and climate variability.

Historically, the Ministry of Health relied on routine health facility reporting through Kenya’s District Health Information System (DHIS2), which aggregates data from more than 17,000 health facilities nationwide. While this system provides extensive monthly data on malnutrition cases, diarrhoeal disease, and low birth weight, traditional forecasting methods based on historical averages have limitations.

These methods assume past trends will continue, making it difficult to anticipate sudden surges triggered by environmental shocks or food insecurity. As a result, interventions often begin after malnutrition rates have already reached critical levels, reducing their effectiveness and increasing the risk of child mortality.

Kenya needed a system capable of forecasting risk early enough to allow preventative action.

Data-Driven Solution

To address this gap, researchers from the University of Southern California (USC) partnered with the Kenya Ministry of Health, Amref Health Africa, and the Microsoft AI

for Good Lab to develop a machine learning model capable of forecasting child malnutrition at the sub-county level.

The model integrates two primary categories of data:

    1. Health System Data such as monthly reports from over 17,000 health facilities, recorded cases of acute malnutrition, and indicators such as diarrhoeal disease treatment and low birth weight.
    2. Satellite Environmental Data such as measurements of gross primary productivity (GPP), which is an indicator of vegetation health and crop growth and environmental signals reflecting drought and agricultural stress.

By combining human health indicators with environmental data, the model captures both immediate clinical trends and broader food system pressures.

Machine Learning Methodology:

Researchers tested multiple modeling approaches and found that gradient boosting machine learning models delivered the strongest predictive performance.

The model was trained to:

detect complex relationships between environmental shocks and malnutrition rates, forecast acute malnutrition prevalence at the sub-county level, and predict risk up to six months in advance.

This multi-variable, data-integrated approach significantly outperformed traditional forecasting techniques.

Implementation and Results

The machine learning model demonstrated strong predictive accuracy:

  • Approximately 89% accuracy for one-month-ahead forecasts
  • Approximately 86% accuracy for six-month-ahead forecasts

Compared to Kenya’s standard “window average” forecasting method, the machine learning model showed substantially improved performance, particularly in regions where malnutrition rates fluctuate unpredictably. The model’s predictive strength was measured using area-under-the-curve (AUC) metrics and it scored 0.86 AUC over six months compared to 0.73 AUC of the traditional baseline model.

This improvement represents a significant advancement in the ability to identify high-risk areas before malnutrition becomes widespread.

To translate these forecasts into actionable policy tools, researchers developed a decision-support dashboard that visualizes predicted malnutrition risk at the sub-county level. The dashboard allows the Kenya Ministry of Health and partner organizations to identify emerging hotspots, allocate therapeutic foods and supplements earlier, deploy community health workers strategically, and eventually adjust nutrition programming proactively.

Outcomes and Impact

  • Early warning capability and improved resource allocation:
    The Kenya case represents a fundamental shift from reactive crisis management to predictive prevention. By forecasting malnutrition risk up to six months in advance, authorities gain critical lead time to mobilize resources before children reach severe wasting stages. While, high-accuracy geographic targeting enables more efficient distribution of ready-to-use therapeutic foods (RUTF), nutrition supplements, and community health interventions.
  • Health system strengthening and scalability:
    The initiative leverages Kenya’s existing DHIS2 infrastructure, demonstrating how national health data systems can be augmented with AI rather than replaced.

    DHIS2 is used in more than 80 low- and middle-income countries, meaning this forecasting framework has strong potential for replication globally.

    Rather than relying solely on emergency appeals after malnutrition rates spike, this approach provides a data-informed mechanism for anticipatory action aiming to mitigate the impact of the hunger crisis.

Conclusion

Kenya’s AI-powered malnutrition forecasting system exemplifies how machine learning can directly support SDG 2: Zero Hunger. By integrating health facility data from over 17,000 reporting centers with satellite-derived environmental indicators, researchers built a predictive model capable of forecasting acute child malnutrition with up to 89% short-term accuracy and 86% six-month accuracy.

This initiative shows that data is not merely descriptive but it can also be predictive, preventative, and life-saving. Through collaboration between academic researchers, government institutions, global health organizations, and AI specialists, Kenya has pioneered a scalable model for anticipatory nutrition response, proving that the ability to act before crisis peaks is as important as the ability to respond after it begins.