以機器學習強化之多因子多狀態模型解析乳癌自然史

Abstract

背景： 瞭解乳癌自然史對於優化篩檢策略與推動精準預防至關重要。乳癌疾病進展通常由正常狀態（free of breast cancer）進入臨床前可檢測期（preclinical detectable phase, PCDP），再進一步發展至臨床期（clinical phase, CP）之動態過程。傳統連續時間馬可夫模型可估計疾病狀態間之轉移速率與平均停留時間，但在處理複雜非線性關係與個體異質性時仍有限制。因此，本研究結合機器學習、多狀態模型與馬可夫神經過程模型，以更完整描述乳癌疾病進展過程。 方法： 本研究分析瑞典雙縣試驗（Swedish Two-County Trial）中Dalarna地區之長期乳癌篩檢追蹤資料，共納入55,145名婦女，追蹤期間自1977年至2022年，建立包含正常狀態、PCDP 及 CP 之三狀態自然史模型。建立羅吉斯迴歸與人工神經網路（artificial neural network, ANN）模型，並以區辨能力與校準度評估模型表現；另使用SHAP值解析特徵對ANN預測結果之貢獻。其次，套用既有文獻參數進行群體層級模擬與個人層級風險分層，並進一步使用本研究資料重新估計傳統指數馬可夫模型與馬可夫神經過程（Markov Neural Process, MNP）模型。 結果： 在正常至PCDP之轉移分析中，僅納入BMI、首次足月生產年齡、乳房緻密度及家族史之傳統危險因子時，模型區辨能力有限，ANN測試集AUC約為0.569；SHAP分析顯示BMI、乳房緻密度及首次足月生產年齡為主要影響因子。於PCDP至CP之轉移分析中，納入腫瘤大小、組織學惡化程度、淋巴結狀態、淋巴血管侵犯及分子亞型等臨床病理與分子標記資訊後，模型區辨能力明顯提升。羅吉斯迴歸模型具有最高測試集AUC約0.848；向後選取後之羅吉斯迴歸模型雖區辨能力略低，但校準表現最佳，且模型較為精簡並具臨床解釋性。ANN模型測試集AUC約為0.836，仍具有一定區辨能力，但校準表現較不理想，因此較適合作為探索非線性風險結構之輔助模型。SHAP分析顯示，腫瘤大小、淋巴結狀態、淋巴血管侵犯及組織學惡化程度為PCDP至CP轉移之重要特徵。 在自然史模擬中，直接套用文獻參數會略低估追蹤至2022年之實際發病情形，當臨床前期發生率調整至約0.0027 時，模擬結果與實際資料較為接近。另外重新估計之傳統指數馬可夫模型可作為穩定之群體平均自然史模型；MNP雖未明顯提升整體模型表現，但可透過個人化生成器矩陣呈現個體轉移速率差異，並提供不確定性推估。 結論： 乳癌自然史不同階段具有不同風險結構。正常至PCDP之早期轉移較難由少數危險因子準確預測；而PCDP至CP之後期進展則與腫瘤病理特徵、淋巴結狀態、淋巴血管侵犯及分子亞型較為密切。就臨床風險預測而言，羅吉斯迴歸向後選取模型在校準度、穩定性與可解釋性上較適合作為主要模型，ANN則可作為探索非線性關係之輔助工具。傳統指數馬可夫模型可提供穩定之群體平均轉移估計，MNP則可補充個人化轉移速率與不確定性推估。本研究整合機器學習、多狀態馬可夫模型、SHAP分析與MNP，提供乳癌自然史分析與個人化風險分層之評估架構，作為未來精準篩檢策略發展之參考。

Background: Understanding the natural history of breast cancer is essential for optimizing screening strategies and advancing precision prevention. Breast cancer progression can be described as a dynamic process in which individuals transition from the free of breast cancer state to the preclinical detectable phase (PCDP), and subsequently to the clinical phase (CP). Traditional continuous-time Markov models can estimate transition rates and mean sojourn times between disease states; however, they remain limited in capturing complex nonlinear relationships and individual heterogeneity. Therefore, this study integrated machine learning, multi-state modeling, and the Markov Neural Process (MNP) model to more comprehensively characterize breast cancer progression. Methods: This study analyzed long-term follow-up data from the Dalarna cohort of the Swedish Two-County Trial, including 55,145 women followed from 1977 to 2022. A three-state natural history model consisting of the free of breast cancer state, PCDP, and CP was constructed. Logistic regression and artificial neural network (ANN) models were developed, and model performance was evaluated using discrimination and calibration measures. SHAP values were further applied to interpret the contribution of features to ANN predictions. In addition, literature-based parameters were used for population-level simulation and individual-level risk stratification. Traditional exponential Markov and Markov Neural Process models were then re-estimated using the study data. Results: In the transition analysis from the free of breast cancer state to PCDP, when only traditional risk factors including BMI, age at first full-term pregnancy, breast density, and family history were included, model discrimination was limited, with the ANN achieving a test AUC of approximately 0.569. SHAP analysis identified BMI, breast density, and age at first full-term pregnancy as the main contributing factors. In the transition analysis from PCDP to CP, model discrimination improved substantially after incorporating clinical pathological and molecular marker information, including tumor size, histological grade, lymph node status, lymphovascular invasion, and molecular subtype. The logistic regression model achieved the highest test AUC of approximately 0.848. Although the backward-selected logistic regression model showed slightly lower discrimination, it had the best calibration performance and was more parsimonious and clinically interpretable. The ANN model achieved a test AUC of approximately 0.836, indicating acceptable discriminative ability; however, its calibration performance was less satisfactory, suggesting that it may be more suitable as a complementary model for exploring nonlinear risk structures. SHAP analysis showed that tumor size, lymph node status, lymphovascular invasion, and histological grade were important features associated with the transition from PCDP to CP. In the natural history simulation, directly applying literature-based parameters slightly underestimated the observed incidence through 2022. After adjusting the preclinical incidence rate to approximately 0.0027, the simulated results became more consistent with the observed data. In addition, the re-estimated traditional exponential Markov model provided a stable population-average natural history model. Although the MNP model did not clearly improve overall model performance, it captured individual differences in transition rates through personalized generator matrices and provided uncertainty estimation. Conclusion: Different stages of breast cancer natural history exhibit distinct risk structures. The early transition from the free of breast cancer state to PCDP was difficult to predict accurately using only a limited set of risk factors, whereas the later transition from PCDP to CP was more closely associated with tumor pathological features, lymph node status, lymphovascular invasion, and molecular subtype. For clinical risk prediction, the backward-selected logistic regression model was more suitable as the primary model in terms of calibration, stability, and interpretability, while the ANN served as a complementary tool for exploring nonlinear relationships. The traditional exponential Markov model provided stable population-average transition estimates, whereas the MNP model offered additional information on personalized transition rates and uncertainty. By integrating machine learning, multi-state Markov modeling, SHAP analysis, and MNP, this study provides a framework for breast cancer natural history analysis and individualized risk stratification, which may serve as a reference for future precision screening strategies.