chevron-right Created with Sketch. back to thoughts

Developing the Mediscout tool with Savics

The availability of effective disease treatment is something that every human being should have access to. Savics, a social IT business, contributes to this matter by leveraging existing technology and its domain knowledge to ensure that even people living in the poorest areas have access to proper medical care. Sagacify was given the honor to support their mission by helping with the development of their Mediscout tool. Mediscout helps with planning, implementing, and monitoring collaborative interventions in the fight against (in this case) tuberculosis in Africa. Through satellite images analysis and machine learning, we helped Savics to become more efficient in identifying and locating people suffering from this disease.

Using Deep Learning to spot the areas

Tuberculosis is still present in many societies, but it is more likely to break out in poor areas. The issue of the higher occurrence rate in poor neighborhoods is due to barriers (financial, geographical, gender-based…) that affect the accessibility of diagnostics and treatment. Savics’ mission is to remove those (financial, geographical, gender-based…) barriers. However, in big cities, they are not yet able to adequately distinguish rich from poor neighborhoods and only have a limited team of doctors to be dispatched in a very large area. Consequently, their field action could be optimized by a more efficient use of their resources, in this case, the adequate distribution of doctors to the different areas.


Sagacify helps them detecting poor neighborhoods in Rwanda by deploying machine learning on satellite images. Dispatching doctors directly to areas with a high probability of containing diseased people greatly intensifies the impact of Savic's actions on the field. Furthermore, it supports the health of people who would otherwise not receive the required treatment.

Let’s get technical

To detect poor neighborhoods in satellite images, we chose to develop a model based on a convolutional neural network architecture (CNN) with transfer learning. The reasoning behind using a CNN is because these models easily detect patterns in images. The layers in a CNN are able to recognize certain features of images and then to combine them, creating more and more sophisticated patterns as we dig deeper into the layers of the network. This behavior was essential to distinguish poor from rich areas on satellite images. We created our model in three steps:

  1. First of all, we downloaded a significant amount of high resolution aerial optical images of African cities with a population of more than 1 million inhabitants and labeled the poor neighborhoods manually in the following manner: whenever we encountered a poor neighborhood, we manually drew bounding boxes around them.
  2. These labeled areas were then used as training data. The nuance here is that from one labeled area, we could derive many more pictures, largely increasing the size of our training set. The CNN was trained to differentiate between “poor”, “normal”, and “rich” areas of these cities.
  3. The final step was to evaluate the performance of the model on a real-life example where the ground truth was known: Rwanda.

Early results and expectations

Recently, they released a small demo powered by Sagacify showcasing what the Mediscout tool can do. Today, the model already achieves an accuracy of 94%, which ensures that Savics' employees in the field will considerably benefit from the model outputs in their everyday actions. Still, we believe that there is still plenty of room for improving its performance:

  • Due to limited training data, the model is not yet able to perfectly understand previously unseen patterns. For instance, the model recognized the graveyard in Kinshasa “La Gombe” as poor instead of “non-residential”.
  • Labeling issues arise due to the practice of cropping images of tiny areas in large labeled bounding boxes. In this case, green areas were only recognized as poor areas.
  • Today, the model predicts three classes only (poor, medium, rich & other). More classes would be useful for a refined and more precise evaluation of the need for medical intervention.

In the future, the model will be improved by feeding it with a larger pool of high-resolution images and executing post-processing on the model predictions. Also extending the model to other African countries will lead to better results, even if they have or have not poor national statistics.

Interested in collaborating with Sagacify to work on ambitious cases? Get in touch with our team!

G e t i n t o u c h search Created with Sketch.