The Application and Comparison of Machine Learning Models for the Prediction of Breast Cancer Prognosis: Retrospective Cohort Study (Preprint)

• DOI: 10.2196/preprints.33440
• Publication Date: 2021-09-14T03:02:25Z
• Authors: Jialong Xiao; Miao Mo; Zezhou Wang; Changming Zhou; Jie Shen; Jing Yuan; Yulian He; Ying Zheng
• Keywords: Brier score; Discriminative model; Data set

<sec> <title>BACKGROUND</title> Over the recent years, machine learning methods have been increasingly explored in cancer prognosis because of appearance improved algorithms. These algorithms can use censored data for modeling, such as support vector machines survival analysis and random forest (RSF). However, it is still debated whether traditional (Cox proportional hazard regression) or learning-based prognostic models better predictive performance. </sec> <title>OBJECTIVE</title> This study aimed to compare performance breast prediction based on Cox regression. <title>METHODS</title> retrospective cohort included all patients diagnosed with subsequently hospitalized Fudan University Shanghai Cancer Center between January 1, 2008, December 31, 2016. After exclusions, a total 22,176 cases 21 features were eligible model development. The set was randomly split into training (15,523 cases, 70%) test (6653 30%) developing 4 predicting overall cancer. discriminative ability evaluated by concordance index (C-index), time-dependent area under curve, D-index; calibration Brier score. <title>RESULTS</title> RSF revealed best among 3-year, 5-year, 10-year curve 0.857, 0.838, 0.781, D-index 7.643 (95% CI 6.542, 8.930) C-index 0.827 0.809, 0.845). statistical difference tested, significantly outperformed Cox-EN (elastic net) (C-index 0.816, 95% 0.796, 0.836; &lt;i&gt;P&lt;/i&gt;=.01), 0.814, 0.794, 0.835; &lt;i&gt;P&lt;/i&gt;=.003), 0.812, 0.793, 0.832; &lt;i&gt;P&lt;/i&gt;&amp;lt;.001). models’ scores very close, ranging from 0.027 0.094 less than 0.1, which meant had good calibration. In context feature importance, elastic net both indicated that TNM staging, neoadjuvant therapy, number lymph node metastases, age, tumor diameter top 5 important A final online tool developed predict <title>CONCLUSIONS</title> slightly other ability, revealing potential method an effective approach building analysis.