A Comparison of Different Models to Recommend Variables for Research Data Requests to German Cancer Registries

Area of Science:

• Oncology
• Data Science
• Bioinformatics

Background:

• Cancer registry data is crucial for research, hypothesis generation, clinical trial recruitment, and evaluating intervention side effects.
• Requesting cancer registry data is often a complex and error-prone process.
• Machine learning (ML) holds potential to streamline data requests for cancer researchers.

Purpose of the Study:

• To compare seven different ML models for variable recommendation in research data requests to German cancer registries.
• To identify the most suitable ML method for assisting researchers in this data request process.

Main Methods:

• Collected data from five German cancer registries, including variable descriptions and analysis details.
• Trained and evaluated seven ML models (neural network, decision trees, random forests, gradient boosting variants) using nested cross-validation.
• Assessed model performance using per-label and per-sample F2 scores.

Main Results:

• A neural network achieved the highest per-label F2 score (approx. 0.3).
• A gradient boosting variant excelled in per-sample F2 score (0.82).
• A random forest variant provided a strong balance between per-sample (0.78) and per-label (0.26) F2 scores, outperforming other models.

Conclusions:

• The random forest model appears best suited for single-model variable recommendation in cancer registry data requests.
• Ensemble models combining neural networks and gradient boosting may offer superior performance but increase complexity.
• Optimizing ML approaches can significantly improve the efficiency and accuracy of accessing vital cancer registry data for research.