Property-Driven Synthetic Data Engineering for Data-Scarce Software Systems: Reflections from the Breast Cancer Domain

2026-07-08 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Saverio D'Amico and Claudio Menghi introduce "property-driven synthetic data engineering" as a critical approach for developing software systems in data-scarce, privacy-sensitive domains like breast cancer treatment. Their work, stemming from a collaboration with oncologists on intraoperative radiotherapy (IORT) data, reveals that synthetic data generation (SDG) shifts the core challenge from data unavailability to defining and validating the specific properties synthetic data must preserve. Using a cleaned dataset of 709 patients and 58 variables, they experimented with SDG techniques including TabDDPM, CTGAN, Gaussian Copula, CopulaGAN, and TVAE, noting TVAE's strong performance in statistical fidelity. However, the authors emphasize that statistical similarity does not equate to clinical plausibility, underscoring the need for multi-stakeholder property definition and validation. They identify key challenges in requirements elicitation under uncertainty, validation without ground truth, and the continuous evolution of synthetic data pipelines.

Key takeaway

For software engineers and AI scientists building data-driven systems in data-scarce, sensitive domains, you must shift from merely generating synthetic data to a property-driven engineering approach. Explicitly define and formalize validity properties—like clinical plausibility, statistical fidelity, and privacy protection—with all stakeholders. Your validation pipelines should continuously check these properties, acknowledging that no single generator is universally optimal and that data cleaning decisions are critical system specifications.

Key insights

Synthetic data engineering requires defining, validating, and evolving stakeholder-specific properties, not just generating data.

Principles

• Synthetic data validity is multi-property.
• No universal best generator exists.
• Data cleaning is an engineering decision.

Method

The proposed method involves stakeholders defining properties, developers operationalizing them as checks, tools comparing generators against checks, and pipelines evolving with changes.

In practice

• Elicit properties from diverse stakeholders.
• Formalize clinical and privacy constraints.
• Monitor pipeline validity as data changes.

Topics

• Property-Driven Synthetic Data
• Data-Scarce Systems
• Breast Cancer
• Intraoperative Radiotherapy
• Clinical Data Privacy
• Software Requirements Engineering

Best for: AI Scientist, Software Engineer, Research Scientist

Related on AIssential

• Synthetic but Not Realistic: The Evaluation Challenge in Generative Modelling for Structured Electronic Medical Records — Current synthetic healthcare data evaluation methods overlook clinical validity, leading to unreliable inferences.
• A Pipeline for Generating Longitudinal Synthetic Clinical Notes Using Large Language Models — A modular LLM-powered pipeline generates privacy-preserving, longitudinal synthetic clinical notes for AI development and evaluation.
• Dozens of AI disease-prediction models were trained on dubious data — AI models for disease prediction are being trained on potentially fabricated public datasets, risking patient harm.
• From virtual experiments to biomedical insight with synthetic data — Bridging the "simulation to reality" gap is crucial for synthetic data to realize its full potential in biomedical AI.
• From Black-Box Confidence to Measurable Trust in Clinical AI: A Framework for Evidence, Supervision, and Staged Autonomy — Clinical AI trust is an engineered system property, not just model accuracy, requiring evidence, supervision, and staged autonomy.
• Possible or Definite? A Benchmark for Evaluating Diagnostic Uncertainty Preservation in Clinical Text — Large language models poorly preserve diagnostic uncertainty in clinical text, posing a significant safety risk.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.