Diagnosis

Context

Children undergoing liver graft have a regular follow-up with specialists. A liver biopsy is harvested roughly every six months. This procedure is very uncomfortable, invasive, and even dangerous (since liver biopsy can even result in the patient’s death). The motivation for regular control is to detect and diagnose early signs of graft rejection or fibrosis. In practice, rejections tend to occur rather early in the post-graft period, while fibrosis will generally develop later on. Symptoms of early rejects are also well detected by experts, in which case, a confirmatory biopsy is almost mandatory. On the other hand, fibrosis is a much more subtle and low-progressing phenomenon, which is not easily detected early enough.

Objective

Our customer, ImmuneHealth, a research center and CRO, wanted to develop an alternative to systematic and repeated biopsy sampling for those patients. The idea was to replace the biopsy by a blood sample, and to extract immunology-related markers from those samples, then have a decision support tool for the fibrosis diagnosis. They collected about 300 biopsies, as well as the corresponding clinical annotations (including the outcome) for all biopsies. They then produced measurements for a series of candidate immuno-markers levels via an in-house multiplex technology applied on the corresponding blood samples.

What we did

  1. Based on the dataset consisting of 300 “rows” (biopsies) and a few tens of “columns” (clinical and immuno features), we identified several “signatures” of features (combining immuno-markers and clinic features) allowing to predict fibrosis, and built diagnosis models on them.
  2. Given that constraints in terms of sensitivity and specificity requirements had been defined, we were able to tune the models accordingly
  3. We incorporated the resulting models into patient management decision trees, fitting in the current clinical practice.

As a result, we designed a solution able to reduce by 25% the number of biopsies on the relevant category of patients, while having a sensitivity of 90%.

Customer’s feedback

“DNAlytics really helped us by fully taking in charge the data processing and analysis. On top of effectively taking in charge the diagnosis modeling, the questions and comments they raised throughout the project led us to better see how our development could fit into clinical practice. They also clearly explained otherwise quite complex modeling concepts.”

Joël Tassignon,
Laboratory Director,
ImmuneHealth.

What if it had to be re-done?

A concern in this project was that of the 300 biopsies collected, many were unusable for the predictive task at hand. Furthermore, samples from specific categories of patients were too few to generate confidence intervals sufficiently narrow at the level of sensitivity and specificity required. If we had the opportunity to write another protocol for a similar project, we would advise a sample collection strategy more in line with the predictive objectives. Also, if the number of candidate biomarkers could have been prospectively reduced, for example based on literature or prior medical expertise, the stringence of the correction of statistical tests for multiple testing would have been reduced.

Do you have to deal with a similar project?

Contact us!