Enhancing clinical outcome predictions through effective sample size evaluation in graph-based digital twin modeling

Area of Science:

• Computational biology and bioinformatics
• Oncology and cancer research
• Health informatics and data science

Background:

• Digital twins offer a novel approach for precision diagnosis, prognosis, and treatment in healthcare.
• SynTwin, a computational method using synthetic data and network science, has shown potential in predicting breast cancer mortality.

Purpose of the Study:

• To validate the SynTwin methodology for generating digital twins using population-level cancer data.
• To assess the predictive accuracy of SynTwin for cancer mortality across diverse cancer types and sample sizes.
• To evaluate the impact of sample size on the predictive performance of digital twin models.

Main Methods:

• Utilized population-level cancer data from the National Cancer Institute's SEER program.
• Generated digital twins using the SynTwin computational methodology, integrating synthetic data and network science.
• Assessed predictive accuracy using nearest network neighbor models, comparing performance with and without digital twins.

Main Results:

• For datasets larger than 10,000 records, incorporating digital twins significantly improved prediction model performance.
• Area Under the Receiver Operating Characteristic (AUROC) curves for cancers like cervix uteri and ovarian cancer reached 0.828–0.884 with digital twins, compared to 0.720–0.858 using real patient data alone.
• Digital twins consistently enhanced AUROCs by at least 0.06 for selected cancers, with reduced performance variance as sample size increased.

Conclusions:

• Network-based digital twins, generated via SynTwin, demonstrate significant benefits for improving cancer mortality prediction accuracy.
• The effectiveness of digital twins is particularly pronounced in larger datasets, enhancing predictive performance beyond models using real patient data alone.
• Effective sample size is a critical consideration for developing robust and accurate predictive models utilizing digital twins.