The validation study performed by our partner on MNM’s AI tool for predicting the response of cancer patients to CDK4/6i treatment is a good example of how advanced the AI in HealthTech is.
The synopsis of the story:
- Cancer patients are given CDK4/6i without an understanding of who will benefit;
- We get genomic and clinical data to build an AI predictive model;
- We run ML/AI to obtain the model;
- We evaluate the model internally and supply it to the partner who validates it.
Let’s get started: we do what the playbook says, but the model is not working. We go back and spend weeks curating the data (cleaning, understanding, labelling, extracting features, etc). Progress is starting to show, and the model is working now (Model 1). However, the results of internal validation are not great.So, we go back and spend weeks improving the playbook (writing new code, debugging tools, benchmarking available tools). It’s going great, the model is improved (Model 2), but we think we can still make it better.We again go back and spend weeks building completely new tools dedicated to the dataset. The result - a model outperforming all previous approaches (Model 3).
In sum, here is where AI in HealthTech (particularly in cancer genomic) is currently:
- The AI can deliver great things (identification of patients who will respond to therapy);
- Most of us think that AI is a magic wand. It is NOT. Most AI companies spend a huge amount of their time curating the data and building novel tools for specific data modalities – so today AI is very human-labor intensive;
- But stay tuned. The AI will soon transition from current incremental growth to an exponential one. You will see that news about new datasets, tools, and insights will give space to solutions to difficult problems. We’re nearly there!