Scholarlink

<p>Accurate and reviewable cost estimation for medical software development is increasingly important in healthcare software environments shaped by governance, documentation, and audit-oriented requirements. Traditional models such as COCOMO and Function Point Analysis often struggle to reflect heterogeneous project conditions, while many machine-learning approaches remain difficult to justify in transparency-sensitive settings. This paper proposes an explainability-aware and constraint-oriented GA–BP framework for medical software cost estimation that combines bounded architecture search with shallow backpropagation neural networks in order to jointly consider predictive competitiveness, architectural simplicity, explanation support, and traceability. Rather than treating explainability as a purely post hoc property, the framework incorporates review-oriented and complexity-aware criteria into model re- tention whenever multiple candidate architectures achieve near-optimal predictive quality. Experiments on a dataset of 127 anonymized, project-level medical software cases show that the proposed framework remains competitive within the studied sample, although it does not outperform the strongest evaluated baseline in raw predictive error. In the present experiment, Linear Regression achieved the lowest mean MAE, while the proposed GA–BP framework clearly outperformed several nonlinear benchmark models and substantially improved over a manually configured BP network. The findings further indicate that the present dataset admits a narrow set of similarly efective shal- low neural architectures, suggesting that the main contribution of the framework lies less in guaranteeing universal predictive dominance and more in formalizing a constraint-aware, audit-oriented model-selection procedure. These results should be interpreted with caution given the limited sample size and the need for broader external valida- tion across medical software subdomains. Overall, the study suggests that prediction quality, explanation support, and audit-oriented model design can be considered together within a unified framework for healthcare software cost estimation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable AI–driven optimization for transparent and auditable decision-making in medical software

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月4日

查看全文
  • Liangyu Li,
  • Zulkefli Mansor,
  • Fatin Filzahti Ismail,
  • Xiaoyan Zhao,
  • Xudong Li,
  • Hao Sun,
  • Xuwei Guo

摘要

Accurate and reviewable cost estimation for medical software development is increasingly important in healthcare software environments shaped by governance, documentation, and audit-oriented requirements. Traditional models such as COCOMO and Function Point Analysis often struggle to reflect heterogeneous project conditions, while many machine-learning approaches remain difficult to justify in transparency-sensitive settings. This paper proposes an explainability-aware and constraint-oriented GA–BP framework for medical software cost estimation that combines bounded architecture search with shallow backpropagation neural networks in order to jointly consider predictive competitiveness, architectural simplicity, explanation support, and traceability. Rather than treating explainability as a purely post hoc property, the framework incorporates review-oriented and complexity-aware criteria into model re- tention whenever multiple candidate architectures achieve near-optimal predictive quality. Experiments on a dataset of 127 anonymized, project-level medical software cases show that the proposed framework remains competitive within the studied sample, although it does not outperform the strongest evaluated baseline in raw predictive error. In the present experiment, Linear Regression achieved the lowest mean MAE, while the proposed GA–BP framework clearly outperformed several nonlinear benchmark models and substantially improved over a manually configured BP network. The findings further indicate that the present dataset admits a narrow set of similarly efective shal- low neural architectures, suggesting that the main contribution of the framework lies less in guaranteeing universal predictive dominance and more in formalizing a constraint-aware, audit-oriented model-selection procedure. These results should be interpreted with caution given the limited sample size and the need for broader external valida- tion across medical software subdomains. Overall, the study suggests that prediction quality, explanation support, and audit-oriented model design can be considered together within a unified framework for healthcare software cost estimation.