多模態深度學習於原發性高醛固酮症之預後預測: 結合臨床與影像資料

Abstract

準確預測術後預後仍然是臨床決策支持中最迫切的挑戰之一。在原發性醛固 酮增多症(PA)中，這一需求尤為重要。PA 是腎上腺分泌過多的醛固酮所致，會 引發高血壓與心血管風險升高。過去認為 PA 在輕中度高血壓患者中較為罕見， 但近期研究估計其盛行率在高血壓人群中已超過 5-10%。PA 通常被分為雙側型 (BPA)與單側型(UPA)。BPA 多以藥物治療為主，而 UPA 則常以單側腎上腺 切除術(uADX)治療。雖然 uADX 可具治癒性，但僅有約 37-43% 的 UPA 患者 能達到完全臨床治癒。因此，可靠的術前預後預測對於指導治療選擇、防止低成功 率患者接受不必要的手術至關重要。然而，目前的預測方法仍有限，受制於臨床生 化特徵、影像資料與患者預後結果之間複雜的交互作用。 本研究提出了一個多模態深度學習框架，用於 PA 的預後預測，整合臨床與影 像資料，並引入新穎的模型架構策略。第一部分為 GAPPA(Graph-basedApproach for Prognosis Prediction in uADX)，其利用圖神經網路(GNN)來捕捉臨床生化特 徵之間的相互依存關係。在一個包含 640 位患者的回溯性資料集中，GAPPA 的表 現顯著優於先進的 TabPFN 與既有機器學習模型，為僅依賴臨床資料的預測建立 了新的基準。在此基礎上，第二部分聚焦於基於影像的預後預測，而這一方面過去 從未將深度學習應用於 PA。本研究開發了一個影像模型 GAMBA (Mamba with Graph-Attention Phase-Fusion for CT-based Prognosis Assessment)，其結合三維卷積 神經網路(3DCNN)、新近的 Mamba 架構，以及基於圖注意力(GAT)的相位融 合機制，用以挖掘多相位三維 CT 掃描中的豐富訊息。這是首次將深度學習應用 於 CT 影像的 PA 預後預測，也是較早期將 Mamba 架構延伸應用至三維 CT 影 像預測的嘗試。在一個包含 213 位具影像資料的患者回溯性資料集中，本研究所 提出之模型表現顯著優於基於放射組學的機器學習方法，並達到與 GAPPA 統計 上相當的預測水平，突顯了多相位影像在此情境下的可行性與臨床價值。 最後，本研究提出 CURE(CenteredUncertainty-AwareRejectEnsemble)以整 合預訓練的臨床與影像模型。不同於傳統的集成方法，CURE 能夠靈活結合來自 不同資料集的模型，從而突破兩種模態必須依賴相同患者群體的限制。透過整合臨床資料集與影像資料集進行訓練與分析，CURE 達成了穩健的預測效能，其平均 F1 分數為 74.7% ± 7.3%，準確率 78.8% ± 5.1，AUROC 為 0.858 ± 0.055，敏感性 72.4% ± 11.2%，特異性 83.8% ± 7.0%。這些結果顯著優於僅用 GAPPA 或僅用影 像模型的表現，同時提升了穩健性、可解釋性與臨床應用的靈活性。 總結來說，本研究建立了首個多模態深度學習框架用於 PA 的預後預測。 臨 床模型 GAPPA 為臨床資料建模樹立了新標準，影像模型 GAMBA 引入了創新的 多相位三維 CT 影像學習方法並將 Mamba 架構拓展此領域，而 CURE 則提供了 一種靈活的不確定性感知策略，以整合異質資料及多模態模型。這些創新共同奠定 了一套具臨床意義的 PA 預後預測方法，不僅有助於提升術前諮詢的準確性與價值， 也展示了臨床與影像資料的多模態整合在推動生物醫學預後建模發展上的潛力。

Accurately predicting postoperative prognosis remains one of the most pressing challenges in clinical decision support. In primary aldosteronism (PA), this need is particularly vital. PA results from excessive aldosterone secretion by the adrenal glands, causing hypertension and heightened cardiovascular risk. Once thought rare in patients with mild-to-moderate hypertension, recent studies estimate its prevalence to exceed 5-10% among hypertensive populations. PA is typically categorized into bilateral PA (BPA) and unilateral PA (UPA). While BPA is usually managed with medications, UPA is commonly treated by unilateral adrenalectomy (uADX). Although uADX can be curative, only 37-43% of UPA patients achieve complete clinical success. Consequently, reliable preoperative prognosis prediction is essential to inform treatment selection, prevent unnecessary surgery in patients with a low likelihood of success. Yet current predictive methods remain limited, constrained by the complex interactions among clinico-biochemical features, imaging data, and postoperative prognosis. This study proposes a multimodal deep learning framework for prognosis prediction in PA, integrating clinical and imaging data with novel architectural strategies. The first component, GAPPA (Graph-based Approach for Prognosis Prediction after uADX), employs graph neural networks (GNNs) to capture interdependencies among clinico-biochemical features. Using a retrospective cohort of 640 patients, GAPPA demonstrated significantly superior performance compared with the state-of-the-art TabPFN and prior machine learning models, thereby establishing a new benchmark for clinical-only prediction. The second component addresses imaging-based prognosis prediction, where deep learning has not previously been applied to PA. A model, GAMBA (Mamba with Graph-Attention Phase-Fusion for CT-based Prognosis Assessment) was developed to combine three-dimensional convolutional neural networks (3DCNN) with the novel Mamba architecture and a phase-fusion mechanism based on graph attention (GAT) to harness the rich information contained in multiphase 3D CT scans. This represents the first application of deep learning to CT imaging for PA prognosis and an early adaptation of the Mamba architecture to 3D CT imaging. In a retrospective cohort of 213 patients with available imaging data, this model outperformed radiomics-based machine learning approaches and achieved performance comparable to GAPPA, underscoring the feasibility and clinical value of multiphase imaging in this setting. Finally, this study proposes CURE (Centered Uncertainty-Aware Reject Ensemble) to integrate the pretrained clinical and imaging models. Unlike conventional ensemble techniques, CURE can flexibly combine models derived from distinct cohorts, thereby overcoming the constraint that both modalities must rely on the same patient population. By leveraging the clinical cohort together with the imaging cohort, CURE achieved robust predictive performance, with F1-score of 74.7% ± 7.3%, accuracy of 78.8% ± 5.1%, AUROC of 0.858 ± 0.055, sensitivity of 72.4% ± 11.2%, and specificity of 83.8% ± 7.0%. These results significantly surpassed those of either GAPPA or GAMBA alone, while enhancing robustness, interpretability, and flexibility for clinical application. In conclusion, this study establishes the first comprehensive multimodal deep learning framework for prognosis prediction in PA. The clinical model GAPPA sets new standards for clinical data modeling, the imaging model GAMBA introduced an innovative multiphase CT-based approach and extended the application of Mamba to this domain, and CURE provided a flexible uncertainty-aware strategy for integrating heterogeneous cohorts and multimodal models. Collectively, these innovations deliver a clinically meaningful framework for PA prognosis prediction, with the potential to improve preoperative counseling, while also demonstrating how multimodal integration of clinical and imaging data can advance prognostic modeling in biomedicine.