80P - Development of a next-generation sequencing diagnostics recommender tool in the framework of the molecular tumor board Freiburg

Background

Revolutions in data-driven personalized healthcare have emerged a crucial component of precision oncology. The choice of which appropriate molecular diagnostic method to initiate – whole-exome sequencing or targeted panel NGS – plays a crucial role for the treatment decision of an oncology patient. With the use of comprehensive patient data along the patient's journey through the MTB process, interdisciplinary molecular tumor boards (MTB) may become more effective in tailoring individualized treatment strategies. As part of an academia-industry-collaboration we designed an approach to reconstruct and support diagnostic decisions in MTBs.

Methods

Historical MTB clinician decisions and objective patient characteristics as possible were used to train in a prediction model to facilitate decision-making. Model design incorporated key Python libraries for solid data cleaning, feature engineering, feature importance analysis, model selection, evaluation, hyperparameter tuning and metrics. We used the LightGBM gradient boosting framework which uses tree-based machine learning principles.

Results

Using MTB data from 2020, we successfully developed a robust prototype NGS recommender, that relies on five key features, including patients age, time from first diagnosis to MTB and number of prior therapies. Validation of the prototype yielded an AUROC score of 0.96 and average precision score of 0.85 for the validation data set.

Conclusions

This study faced several limitations, including an imbalance in the representation of the two NGS methods in the dataset and inaccessible features such as molecular pathological pre-findings presented in plain text. However, despite these constraints, the development of the prototype NGS recommender showed promising performance and represents a significant step towards a decision support tool for molecular diagnostics. Moving forward, validation of the recommender on updated and then external datasets remain a crucial next step. In addition, exploration of follow-up patient data and resolution of relevant implementation logistics will be essential for successful integration and use in the clinical setting, providing an opportunity for iterative improvement.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.