本研究探討急性腦中風出院功能預後之機器學習預測，旨在維持預測效能之同時，降低模型於臨床實務之導入門檻。現行預後評分與機器學習模型多仰賴高維度資料，且常以中風後 30 至 90 天的預後結果為預測目標，難以完全契合臨床團隊於住院期間即需規劃出院照護資源之決策時效需求。為此，本研究以出院時改良式雷氏量表（modified Rankin Scale， mRS）為基準，將預後分為「良好（0-2分）」與「不良（3-6分）」進行二元分類預測，並將極端特徵精簡與決策解釋性作為核心設計原則，建構兼具精準度與實務可行性之預後模型。 研究回溯性納入聯新國際醫院 2010 至 2022 年之中風登錄資料，經系統化變項篩選與特徵工程，最終確立 7，590 筆住院事件與 96 個候選特徵。為模擬真實部署情境，本研究採時間分割方式建置獨立測試集，並運用梯度提升樹（GBDT）演算法建構分類模型。針對「早期預測（入院 72 小時內）」與「全期預測（自入院至出院）」兩種情境，本研究建立以排列重要度（permutation importance）為基礎之多階段特徵精簡流程。透過幾何手肘法與累積佔比法進行雙軌初步篩選，復於核心模型中導入前向逐步選取策略，精準界定效能高原期之截斷點。 結果顯示，臨床數據中存在明確之「資訊飽和」現象：早期與全期情境分別僅需 11 項與 17 項特徵構成之「最小充分預測特徵集」，其預測表現即穩定達到 AUC 值 0.83 與 0.85，，且 PR-AUC 亦維持良好水準，與完整特徵模型相比未見實質效能衰退。進一步運用逐群組剔除法（LOGO）進行消融分析發現，預後主導因子隨病程產生重心轉移：早期高度仰賴初始神經學嚴重度與基線特徵；全期則顯著受住院期間之醫療處置、併發症管理與生理穩定度影響。此外，反事實模擬揭示了預後風險之高度非對稱性：低風險個案對單一不利擾動極為敏感，而高風險個案往往需仰賴多面向照護組合方能產生協同之預後翻轉效果。在極限救援情境中，早期模型之預後可挽救率高達 45.9%，，至全期階段則大幅縮減至 23.5%，，具量量化了療療金時時間與早期入入之重要性。結結而，，本研究提出一套以精簡特徵集為基礎之急性中風預後評估架構，透過反事實模擬將風險預測延伸為決策輔助工具。其不僅有效降低臨床應用門檻，更能協助醫療團隊於病程早期辨識關鍵入入因子，並於後期務實調整預後溝通與長期照護規劃。;This study investigates machine learning-based prediction of functional outcomes at discharge for acute stroke， aiming to lower the barriers to clinical implementation while maintaining predictive performance. Current prognostic scores and machine learning models often rely on high-dimensional data and typically focus on medium-term prognostic outcomes at 30 to 90 days. This focus fails to fully meet the time-sensitive decision-making needs of clinical teams， who must plan post-discharge care resources during the patient′s hospitalization. To address this gap， this study sets the prediction of functional outcomes at discharge－binarized into good (mRS 0-2) versus poor (mRS 3-6) outcomes－as the target. This retrospective study included stroke registry data from Lien Shin International Hospital between 2010 and 2022. Following systematic variable screening and feature engineering， a final dataset comprising 7，590 inpatient episodes and 96 candidate features was established. To simulate a real-world deployment scenario， a temporal split method was utilized to create an independent test set， and Gradient Boosting Decision Tree (GBDT) algorithms were employed to construct classification models. Targeting two scenarios－"early prediction (within 72 hours of admission)" and "full-course prediction (from admission to discharge)"－this study developed a multi-stage feature simplification pipeline based on permutation importance. A dual-track preliminary screening utilizing the geometric elbow method and cumulative importance thresholding was conducted， followed by the introduction of a forward stepwise selection strategy into the core models to precisely define the cutoff point at the performance plateau. The results reveal a clear phenomenon of "information saturation" within the clinical data: the early and full-course scenarios required a "minimal sufficient predictive feature set" of only 11 and 17 features， respectively， to stably achieve AUCs of 0.83 and 0.85. PR-AUC values also remained robust， demonstrating no substantial performance degradation compared to the full-feature models. Furthermore， Leave-One-Group-Out (LOGO) ablation analysis demonstrated a shift in the dominant prognostic factors along the disease course: early prediction relied heavily on initial neurological severity and baseline characteristics， whereas full-course prediction was significantly influenced by medical interventions， complication management， and physiological stability during hospitalization. Additionally， counterfactual simulation revealed a high asymmetry in prognostic risk: low-risk individuals were extremely sensitive to a single unfavorable perturbation， while high-risk individuals often required multifaceted care bundles to produce a synergistic outcome-reversal effect. In the extreme rescue scenario， the prognostic reversibility rate in the early model reached 45.9%， but this proportion shrank drastically to 23.5% in the full-course stage， thereby objectively quantifying the therapeutic golden window and the necessity of early intervention. In conclusion， this study proposes an acute stroke prognostic evaluation framework based on parsimonious feature sets， extending risk prediction into a decision-support tool through counterfactual simulation. This framework not only effectively lowers the threshold for clinical application but also assists medical teams in identifying key intervention targets in the early stages and pragmatically adjusting prognostic communication and long-term care planning in the later stages.
