Examples of Akai Kaeru at work in the Biotech industry

Akai Kaeru’s Explainable AI for Biotech

Our software can quickly identify the conditions under which features become important and then uncover the causal relations that connect them. First, our pattern mining engine identifies groups of data items with shared characteristics. These patterns are then passed on to our easy-to-use interactive Visual Pattern Explorer where analysts can inspect and act upon them. Use cases in Biotech include:

Furthermore, our correlation miner will help you determine:

Case Study:

Identify Salient Factors in Friedreich’s Ataxia

We took a data set comprising history data of subjects with Friedreich’s ataxia. Our purpose was to identify factors that could explain variations in the progression of the disease as measured by the modified Friedreich’s Ataxia Rating Scale (mFARS). The mFARS scale is an exam-based neurological test in which a higher score indicates more advanced progression. The goal of the study was to specifically identify factors that would indicate a higher than average one year change in the mFARS scale (denoted as cMFARS). These factors could then be used during a clinical trial to select subjects, which are likely to progress at a higher rate.

We used AK Miner to identify groups of subjects for which disease progression was unusually high or low. These groups were statistically robust and defined by simple criteria (e.g., low Age at Onset or Low Baseline mFARS). We then switched to AK Visual Pattern Explorer to view and further explore this information, as seen in the figures below.

Akai Kaeru’s XAI Dashboard:

See It All in One Place

The Feature Importance Chart (bottom left) reveals that Age is one of the most important predictor variables for cMFARS. Clicking the Age bar in this chart populates the Group Bubble Chart (top left). The y-axis is the target variable, cMFAR and the x-axis is the selected predictor variable, Age. Each bubble represents a group of subjects who have similar characteristics and for which cMFARS is unusually high (shaded green) or low (red). The opacity of a bubble reveals how important the x-axis feature (here Age) is in defining the bubble’s group. In this specific example, we can observe a strong trend in which lower age is associated with higher disease progression. In this case, the opacity of the groups tells us that age is only really important at its low values (i.e., there is an accelerated disease progression in children).

The user can click on a bubble for further analysis of its respective group. This adds a dark outline to the bubble. The Group Summary on the top right panel then shows the corresponding detailed information. We learn that the selected group is described by low Age (i.e., age <= 15) and low Upper Limb Coordination (ULC) score (i.e., ULC <=13). The text in the right panel tells us that this is a statistically significant finding.

In the Probability Histogram panel (bottom center of the dashboard) we can see the distribution of cMFARS for both the selected group (green) and the full set of subjects (gray). The bottom right panel shows the summary statistics for the selected group. Finally, the Feature Importance panel on the bottom left shows the importance of the different features both globally (gray) as well as the important features for the selected group (blue). Here we can see that although there are features that are more important globally, Age (as well as Education Level, the top bar) is more important for the selected group.

This is just one example of the many patterns of interest found using Akai Kaeru’s pattern mining and pattern browsing interface. Some other useful features not covered here include exploring alternative explanations for a specific group, correlation mining (i.e., finding groups of subjects for which two attributes become highly correlated), and statistically controlling for the effects of confounding variables.