Currently, I am working on a project to build an anemia detection model using 14 features, such as hemoglobin level, white blood cell count, and others. The model’s output is the type of anemia, classified into 9 possible labels. I have developed a neural network model that achieves an accuracy of 94% and a loss of 0.5 on the test dataset. How can I further improve my model? Here’s the link to the dataset: link
Use other metrics such as recall, precision and F1. Also, if the model doesn’t perform that well, maybe choose more (other) features that might be more relevant to what you are trying to predict. The usual advice of model architecture revision may also be appropriate!
1 Like
This paper uses images of the hand as the data set, but you might find their approach to analysis and measurement stimulating. Some of the model algorithms discussed reached above 99% and 96% was the worst of several.
At first glance, it seems they followed all 3 of @gent.spah ’s recommendations; additional measures, different features, and different architectures.
Let us know what you discover?
1 Like
Great article @ai_curious. Thank you!
1 Like
Here is a link to another paper you might find interesting. It further supports @gent.spah ’s recommendation regarding metrics.
On evaluation metrics for medical applications of artificial intelligence
1 Like