Machine learning for medical risk assessment
Alongside heart disease and cancer, diabetes is one of the most common serious health issues today.
We turned to ML in order to improve the accuracy in predicting increased risk for diabetes. After training and fine tuning we can look closer at the explainability of the models using explainable AI
500'000+
people in Sweden have diabetes and more than 400 million worldwide with a rapid increase in developing countries
1 / 3
are unaware that they have diabetes and 9/10 among those with increased risk do not know about it
<40%
of people with increased risk of diabetes are identified during a standardized health check
Health issues
for diabetes include blindness, kidney failure, heart disease, stroke and loss of limbs
Using XAI to unbox the models
Global and regional explainability
- XAI can provide different stakeholders with an intuitive understanding of predictive machine learning models.
- The ‘global’ view provides an understanding of the model across all data points, such as the overall feature importances.
- Conversely ‘regional’ feature importances can differ from the global average. An example is showed for ‘Age’ where the importance of the feature is more significant for larger deviations from the mean. Various regional views can help identify and quantify biases in the model.
Local explainability
- Looking closer at individual observations, one can get a better understanding of the contributions for a set of feature values.
- An example observation shows an increased risk for diabetes due to ‘Age’ and ‘Waist circumference’. Note that the ‘HbA1c’ level, often used as the single risk measure in standardized health screenings, would in this case have failed to capture the overall risk.