肝細胞癌（Hepatocellular Carcinoma， HCC）為全球癌症相關死亡之第三大主因。對於確診時即不適合手術切除或肝臟移植之患者（約占全部病例六至七成），局部治療（Local-Regional Therapy， LRT）為主要治療選項；惟治療後之早期進展（Early Progression， EP）發生率達 25% 至 45%，代表腫瘤侵襲性較高或消融不完全，顯著縮短存活。臨床上之預後判斷面臨三項瓶頸：同一分期或肝功能等級內病人預後仍高度異質、缺乏可再細分之客觀工具；預後由腫瘤生物學與肝功能儲備等多面向因子共同決定，單一資料面向不足且缺乏整合方法；既有模型多為複雜評分或機器學習黑箱，難以於床邊解釋與使用。 本研究以 95 位於臺大醫院接受局部治療之肝細胞癌患者為對象（早期進展 37 人、死亡 15 人），整合臨床與生化指標及血漿游離 DNA（cell-free DNA， cfDNA）甲基化標記，建立非監督式整合分群之預後模型。研究先檢驗監督式預測之可行性，發現其曲線下面積（Area Under the Curve， AUC）達上限約 0.7 且於隨機切分下高度不穩定；遂改採非監督式整合分群為主軸。為避免單一方法或參數造成偏誤，本研究系統性比較 120 種配置（2 種基因封頂上限 × 4 種變數組合 × 3 種群數 × 5 種分群演算法），並依每變數事件數（events-per-variable， EPV）確立群數 k= 3，以存活顯著性與綜合分數進行多準則篩選。 結果顯示資料自然分為三個亞型，依預後遞減命名為 C1 肝功能良好型、C2 肝炎活化型與 C3 肝功能異常型（人數分別為 54、17、24）。三亞型於整體存活（Overall Survival， OS）與無惡化存活（Progression-Free Survival， PFS）皆顯著分層（整體存活 p= 0.004），主模型（高斯混合模型）與副模型（共識分群）有 92.6% 病人分至相同群、ARI= 0.78。模態消融分析顯示整合生化與基因之一致性指數（C-index）最高（整體存活 0.767），移除任一模態後即降至接近隨機（0.52 至 0.53）。以巢狀 Cox 迴歸與概似比檢定（Likelihood Ratio Test， LRT）可知，分群於校正 Child-Pugh 分級後仍顯著（p= 0.017 與 0.027），而反向不成立；最差之 C3 對最佳之 C1 整體存活風險比（Hazard Ratio， HR）達 7.84。最終以分類與迴歸樹（CART）將分群規則萃取為僅需轉胺酶、白蛋白與球蛋白比值及膽紅素三項常規抽血之決策樹，留一交叉驗證（Leave-One-Out Cross-Validation， LOOCV）重現率達 85%。 本研究證實，整合生化與基因之非監督式分群可提供可重現、且優於 Child-Pugh 分級之整體存活預後分層（一致性指數 0.76），並可萃取為床邊可手算之三變數決策樹。cfDNA 甲基化於靜態單一時點之分群中未顯現獨立預後力，其價值主要展現於分群形成階段。本研究為小樣本臨床預後場景提供「整合勝於單一、方法學嚴謹且臨床可轉譯」之實務範例。;Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality worldwide. For patients ineligible for surgical resection or transplantation， local-regional therapy (LRT) is the mainstay of treatment; however， early progression (EP) occurs in 25% to 45% of cases after LRT and markedly shortens survival. Prognostic assessment faces three bottlenecks: substantial residual heterogeneity within the same stage or liver-function grade， with no objective tool for finer risk stratification; the multifactorial nature of prognosis， which no single data modality can capture and for which an integrative method is lacking; and the limited bedside usability of existing complex or black-box models. In 95 HCC patients who received LRT at National Taiwan University Hospital (37 with EP， 15 deaths)， this study integrated clinical and biochemical indicators with plasma cell-free DNA (cfDNA) methylation markers to build an unsupervised integrated-clustering prognostic model. A supervised prediction attempt reached a ceiling AUC of about 0.7 and was unstable under random splits， motivating a pivot to unsupervised clustering. To avoid bias from any single method or parameter， 120 configurations were systematically compared (2 gene caps x 4 variable combinations x 3 cluster numbers x 5 clustering algorithms). The cluster number k= 3 was fixed by the events-per-variable (EPV) criterion， and configurations were ranked by survival significance and a composite score. Three subtypes emerged， named by descending prognosis as C1 (well-compensated)， C2 (active hepatitis)， and C3 (liver dysfunction)， with 54， 17， and 24 patients， respectively. The three subtypes were significantly stratified for both overall survival (OS) and progression-free survival (PFS) (OS p= 0.004); the main model (Gaussian mixture model) and the secondary model (consensus clustering) assigned 92.6% of patients to identical clusters (ARI= 0.78). Modality ablation showed that the integrated model achieved the highest concordance index (OS C-index 0.767)， whereas removing either modality reduced discrimination to near chance (0.52 to 0.53). In nested Cox regression with likelihood-ratio testing， clustering remained significant after adjusting for Child-Pugh class (p= 0.017 and 0.027)， but not vice versa; the OS hazard ratio of C3 versus C1 reached 7.84. A classification and regression tree (CART) distilled the clustering rule into three routine blood tests (transaminase， albumin/globulin ratio， and bilirubin)， with 85% leave-one-out cross-validation (LOOCV) reproducibility. This study demonstrates that integrated unsupervised clustering of biochemical and genetic data provides reproducible OS stratification (C-index 0.76) superior to Child-Pugh grading， and can be distilled into a hand-computable three-variable decision tree. cfDNA methylation showed no independent prognostic power in static single-timepoint clustering， with its value mainly expressed during cluster formation.
