DSpace collection: 博碩士論文

結合全基因組關聯分析與機器學習之憂鬱症基因體風險預測流程建構－以臺灣人體生物資料庫為例;結合全基因組關聯分析與機器學習之憂鬱症基因體風險預測流程建構－以臺灣人體生物資料庫為例

Tue, 20 Oct 2026 06:22:30 GMT

title: 結合全基因組關聯分析與機器學習之憂鬱症基因體風險預測流程建構－以臺灣人體生物資料庫為例;結合全基因組關聯分析與機器學習之憂鬱症基因體風險預測流程建構－以臺灣人體生物資料庫為例 abstract: 憂鬱症（Major Depressive Disorder， MDD）為常見且疾病負擔沉重之精神疾病，其形成涉及遺傳、環境、生活型態與心理社會因素，並具有高度異質性與多基因特性。現有憂鬱症基因體研究多以歐美大型族群為主，臺灣族群之基因體風險分類研究仍相對有限。因此，本研究以臺灣人體生物資料庫（Taiwan Biobank， TWB）為基礎，建立一套結合全基因組關聯分析（Genome-Wide Association Study， GWAS）與機器學習之憂鬱症病例／健康對照分類流程，並評估其於臺灣族群中的方法學可行性。本研究建立憂鬱症病例組與嚴格健康對照組，並進行樣本與單核苷酸多型性（Single Nucleotide Polymorphism， SNP）品質控管、連鎖不平衡剪枝、親緣關係推估與基因關係矩陣（Genomic Relationship Matrix， GRM）建構。為降低親屬跨越資料集與測試集資訊提前進入模型所造成之效能高估，本研究採用親屬群組切分（family-aware split），使同一親屬群組不會同時出現在訓練集與測試集，並僅於訓練集執行 mixed model GWAS、候選 SNP 篩選與特徵建構。訓練集 GWAS 於 P < 1 × 10⁻² 門檻下取得 5，088 個候選 SNP，作為後續分析之候選母集合。在特徵建構方面，本研究整合 SNP 訊號強度、效果方向一致性、交叉驗證穩定性與重抽樣穩定性，將 5，088 個候選 SNP 分為 stable biological core 1，196 個、extended support layer 1，804 個與 polygenic background 2，088 個，並建立 SNP100、SNP500、SNP1000、SNP2000、SNP3000 與 All-SNP 等不同 SNP pool。模型建構比較 raw dosage 與 score vector representation、age_sex 與 cov_core_family 協變量模式，以及 Logistic Regression、Random Forest、XGBoost、LightGBM、CatBoost、Soft Voting 與 Stacking 等模型。模型效能以 AUC、Accuracy、Balanced Accuracy、Sensitivity、Specificity、F1-score 與 Matthews Correlation Coefficient（MCC）等指標評估。研究結果顯示，All-SNP（5，088 SNP）搭配 raw dosage、cov_core_family 與 Stacking 為內部保留測試集中整體表現最佳之組合，其 AUC 為 0.8987、Accuracy 為 0.8337、Balanced Accuracy 為 0.8225、Sensitivity 為 0.8024、Specificity 為 0.8426、F1-score 為 0.6798、MCC 為 0.5834。SNP3000 搭配 raw dosage、cov_core_family 與 Stacking 之 AUC 為 0.8571、Balanced Accuracy 為 0.7809，顯示在減少候選 SNP 特徵後仍可保留相當程度之分類能力。於固定 SNP3000、Logistic Regression、age_sex 與相同 train-OOF threshold 規則下，10-bin score vector 將特徵數由 3，002 降至 32，而 Balanced Accuracy 由 0.7661 降至 0.7559，顯示其具有大幅降維之可行性。綜合而言，本研究建立一套由親屬群組切分、train-only GRM mixed model GWAS、候選 SNP 證據分層、SNP representation、協變量整合、多模型比較與 SNP-to-gene 註解所構成之可追溯流程。研究結果支持此流程於臺灣族群憂鬱症病例與嚴格健康對照分類之內部方法學可行性；惟目前結果仍屬研究性風險分類，尚未完成外部獨立驗證，不宜直接解讀為個人未來發病預測或臨床診斷工具。;Major depressive disorder (MDD) is a heterogeneous and polygenic psychiatric disorder influenced by genetic， environmental， lifestyle， and psychosocial factors. Most genomic studies of depression have been conducted in populations of European ancestry， whereas genomic risk classification studies focused on the Taiwanese population remain limited. This study therefore developed a genomic case-control classification pipeline for MDD by integrating genome-wide association study (GWAS) and machine learning using data from the TWB (TWB). MDD cases and strict healthy controls were established from TWB phenotype data. Genotype quality control， linkage disequilibrium pruning， kinship estimation， and genomic relationship matrix (GRM) construction were performed. A family-aware split was applied so that genetically related individuals were not assigned to both training and testing sets. Mixed model GWAS， candidate SNP selection， feature ranking， and preprocessing were conducted using the training set only. A total of 5，088 candidate SNPs were identified at GWAS P < 1 × 10⁻² and used as the candidate feature universe. The 5，088 candidate SNPs were further stratified into a stable biological core (1，196 SNPs)， an extended support layer (1，804 SNPs)， and a polygenic background (2，088 SNPs) based on signal strength， effect-direction consistency， cross-validation stability， and resampling stability. SNP100， SNP500， SNP1000， SNP2000， SNP3000， and All-SNP pools were evaluated. Raw SNP dosage and score-vector representations， age_sex and cov_core_family covariate settings， and seven classifiers (Logistic Regression， Random Forest， XGBoost， LightGBM， CatBoost， Soft Voting， and Stacking) were compared. The All-SNP model using 5，088 raw SNP features， cov_core_family covariates， and Stacking achieved the highest performance in the internal hold-out test set， with an AUC of 0.8987， Accuracy of 0.8337， Balanced Accuracy of 0.8225， Sensitivity of 0.8024， Specificity of 0.8426， F1-score of 0.6798， and MCC of 0.5834. The SNP3000 raw-dosage model with cov_core_family and Stacking achieved an AUC of 0.8571 and Balanced Accuracy of 0.7809. Under a fixed SNP3000， Logistic Regression， age_sex， and train-OOF threshold setting， 10-bin score vector reduced the feature count from 3，002 to 32 while Balanced Accuracy decreased from 0.7661 to 0.7559. In summary， this study developed a traceable genomic classification pipeline integrating family-aware splitting， train-only GRM mixed model GWAS， candidate SNP evidence stratification， SNP representation， covariate integration， multi-model comparison， and SNP-to-gene annotation. The findings support the internal methodological feasibility of classifying MDD cases and strict healthy controls in a Taiwanese population. External validation is still required， and the current model should not be interpreted as a clinical diagnostic tool or a validated predictor of future disease onset.

應用人工智慧建模於高危險嬰幼兒及早產兒早期動作影像之兒童動作發展預測;Application of Artificial Intelligence Modeling to Early Motor Video Analysis for the Prediction of Motor Development in High-Risk and Preterm Infants

Tue, 20 Oct 2026 06:22:29 GMT

title: 應用人工智慧建模於高危險嬰幼兒及早產兒早期動作影像之兒童動作發展預測;Application of Artificial Intelligence Modeling to Early Motor Video Analysis for the Prediction of Motor Development in High-Risk and Preterm Infants abstract: 腦性麻痺與動作發展異常為高危險嬰幼兒及早產兒常見之神經發展問題。由於其腦部發育尚未成熟，較易受到缺氧、感染及腦部損傷等因素影響，進而導致動作與姿勢控制異常，並可能合併感覺、認知、語言及行為等問題，對日常生活功能與社會參與造成長期影響，亦增加家庭與醫療體系之照護負擔。然而，傳統臨床診斷多須待明顯動作異常出現後方能確認，往往錯失神經可塑性最高的早期介入時機，因此，建立有效的早期辨識方法具有重要臨床意義。嬰幼兒早期自發性動作可反映中樞神經系統之功能完整性，其中動作的流暢度、協調性與變化性，與後續神經發展預後密切相關。一般動作評估（General Movement Assessment， GMA）雖已被證實具良好預測效度，但其高度仰賴專業訓練與臨床經驗，限制了大規模推廣與臨床實務應用。基於此，本研究結合嬰幼兒早期自發性動作影像與人工智慧分析技術，透過骨架姿態特徵擷取與機器學習建模，探討早期動作表現與後續動作發展結果之關聯性，期望建立人工智慧輔助之分析方法，作為高危險嬰幼兒及早產兒早期發展風險評估之依據。本研究共納入 9 名高危險嬰幼兒及早產兒，蒐集其新生兒期自發性動作影片，並追蹤矯正年齡 6 個月、12 個月及 24 個月之粗大動作發展結果。利用 OpenPose 進行骨架姿態擷取與特徵分析，結合出生週數、出生體重及性別等臨床資料，建立人工智慧預測模型，以評估早期動作表現對後續發展結果之預測能力。研究結果顯示，部分預測模型最高準確率可達 93.33%，顯示嬰兒早期自發性動作具有預測後續動作發展之潛力。其中，角度位移特徵（AngDisp）於短期預測表現最佳，而相對關節方向特徵（Rel_JO）於中長期預測展現較佳表現，顯示動作協調性、姿勢控制及肢體間相對關係可能為反映神經發展的重要指標。此外，出生體重對模型表現具有一定輔助效果，但動作特徵對預測效能之影響較臨床變數更為顯著。綜合而言，本研究證實高危險嬰幼兒及早產兒之早期自發性動作與後續發展結果具有關聯性，人工智慧結合動作影像分析可提供客觀之預測依據，具作為臨床早期篩檢與追蹤輔助工具之潛力。未來仍需擴大樣本數並進行多中心驗證，以提升模型之穩定性與臨床應用價值。;Cerebral palsy and motor developmental abnormalities are common neurodevelopmental disorders in high-risk and preterm infants. Due to the immaturity of brain development， these infants are more susceptible to hypoxia， infection， and brain injury， which may lead to impairments in motor and postural control. In addition， sensory， cognitive， language， and behavioral deficits may coexist， resulting in long-term limitations in daily functioning and social participation and imposing a substantial burden on families and healthcare systems. However， conventional clinical diagnosis is often delayed until obvious motor abnormalities become apparent， thereby missing the critical period of maximal neuroplasticity for early intervention. Therefore， establishing an effective method for early identification is of considerable clinical importance. Early spontaneous movements reflect the functional integrity of the central nervous system， and their fluency， coordination， and variability are closely associated with later neurodevelopmental outcomes. Although General Movement Assessment (GMA) has been demonstrated to possess good predictive validity， its implementation relies heavily on specialized training and clinical experience， limiting its application in large-scale screening and routine clinical practice. Therefore， this study aimed to investigate the relationship between early spontaneous motor performance and subsequent developmental outcomes by integrating spontaneous movement video analysis with artificial intelligence techniques. A total of nine high-risk and preterm infants were enrolled in this study. Spontaneous movement videos recorded during the neonatal period were collected， and gross motor developmental outcomes were followed up at corrected ages of 6， 12， and 24 months. Skeletal pose information was extracted using OpenPose， followed by feature extraction and data preprocessing. Machine learning models were constructed by incorporating kinematic features and clinical variables， including gestational age， birth weight， and sex. The results demonstrated that several prediction tasks achieved a maximum accuracy of 93.33%， indicating the potential of early spontaneous movements for predicting later motor development. Among the extracted features， angular displacement (AngDisp) exhibited superior performance in short-term prediction， whereas relative joint orientation (Rel_JO) demonstrated greater robustness in mid- and long-term prediction. These findings suggest that movement coordination， postural control， and inter-limb relationships may serve as important indicators of neurodevelopmental outcomes. In addition， gestational age contributed positively to model accuracy， whereas kinematic characteristics exerted a greater influence on predictive performance than clinical variables. In conclusion， the findings of this study indicate that early spontaneous movements in high-risk and preterm infants are closely associated with subsequent motor developmental outcomes. Artificial intelligence-based motion analysis provides an objective and quantitative approach for early developmental risk assessment and has the potential to serve as an adjunctive tool for clinical screening and follow-up. Future studies with larger sample sizes and multicenter validation are warranted to further improve model robustness and enhance its clinical applicability. Keywords: high-risk infants， preterm infants， cerebral palsy， General Movement Assessment (GMA)， artificial intelligence modeling， early motor image analysis

跨物種EEG資料於睡眠與清醒分類之深度學習可行性研究:從小鼠到人類;跨物種EEG資料於睡眠與清醒分類之深度學習可行性研究:從小鼠到人類

Tue, 20 Oct 2026 06:22:26 GMT

title: 跨物種EEG資料於睡眠與清醒分類之深度學習可行性研究:從小鼠到人類;跨物種EEG資料於睡眠與清醒分類之深度學習可行性研究:從小鼠到人類 abstract: 睡眠分期為睡眠醫學與神經科學研究中的重要課題，傳統睡眠分期通常依賴人工判讀多導睡眠檢查（Polysomnography， PSG）資料，不僅耗時且具有主觀性。隨著深度學習技術的發展，自動化睡眠分類方法逐漸成為重要研究方向。本研究利用連續小波轉換（Continuous Wavelet Transform， CWT）將腦波訊號（Electroencephalography， EEG）轉換為時頻圖表示，並結合卷積神經網路（Convolutional Neural Network， CNN）進行睡眠分類分析。本研究分別針對小鼠 EEG 與人類 EEG 睡眠資料建立分類模型。在小鼠 EEG 實驗中，以清醒（Awake）與睡眠（Sleep）兩種狀態進行二分類，並採用以個體為分群之交叉驗證方式評估模型表現。實驗結果顯示，小鼠 EEG 二分類模型在交叉驗證與外部測試資料上皆可達到約 99% 的分類準確率，顯示模型能有效辨識小鼠清醒與睡眠狀態之腦波特徵。在人類 EEG 睡眠分期實驗中，本研究使用 Sleep-EDF Database Expanded 資料集進行六分類睡眠階段辨識，並採用五折交叉驗證評估模型效能。實驗結果顯示，人類 EEG 六分類模型平均分類準確率約為 76%。雖然多分類睡眠分期任務較二分類具有更高複雜度，但結果仍顯示以 CNN 結合 EEG 時頻特徵進行睡眠階段辨識具有良好可行性。此外，進一步導入跨物種轉移學習（transfer learning）策略，將以人類 EEG 訓練之預訓練模型應用於小鼠 EEG 分類任務中。實驗結果顯示，跨物種轉移學習模型於小鼠 Awake/Sleep 二分類任務中之平均分類準確率約為 98.4%，相較於由頭訓練之小鼠 EEG 基準模型（約 99.8%）略低。然而，轉移學習模型仍維持接近 98% 以上之分類準確率，顯示人類 EEG 所學習之時頻特徵並未因物種差異而完全失效，在一定程度上仍具有跨物種之可遷移性。此結果亦支持利用人類 EEG 預訓練模型作為小鼠 EEG 分類之特徵提取基礎具有可行性。綜合而言，本研究驗證了深度學習結合 EEG 時頻表示於睡眠分類任務中的有效性，並初步探討跨物種特徵遷移之可行性。研究成果顯示，卷積神經網路不僅能有效擷取腦波特徵，亦具備跨物種應用之潛在能力，為未來 EEG 睡眠分析與神經科學研究提供重要參考。;Sleep staging is an important topic in sleep medicine and neuroscience research. Traditional sleep staging generally relies on manual scoring of polysomnography (PSG) recordings， which is both time-consuming and subjective. With the rapid development of deep learning techniques， automated sleep classification methods have become an increasingly important research direction. In this study， electroencephalography (EEG) signals were transformed into time-frequency representations using Continuous Wavelet Transform (CWT)， and Convolutional Neural Networks (CNNs) were employed for sleep classification analysis. Classification models were developed for both mouse EEG data and human EEG sleep datasets. In the mouse EEG experiment， a binary classification task was performed to distinguish between Awake and Sleep states. A subject-based cross-validation strategy was adopted to evaluate model performance. The experimental results demonstrated that the mouse EEG classification model achieved approximately 99% classification accuracy in both cross-validation and external testing datasets， indicating that the proposed model could effectively identify EEG characteristics associated with wakefulness and sleep in mice. For human EEG sleep staging， the Sleep-EDF Database Expanded dataset was utilized to perform six-class sleep stage classification. Five-fold cross-validation was employed to evaluate model performance. The experimental results showed that the human EEG classification model achieved an average accuracy of approximately 76%. Although multi-class sleep staging is considerably more complex than binary classification， the results indicate that CNNs combined with EEG time-frequency features provide a feasible approach for automatic sleep stage recognition. Furthermore， a cross-species transfer learning strategy was investigated by applying a pre-trained model trained on human EEG data to a mouse EEG classification task. The experimental results demonstrated that the transfer learning model achieved an average classification accuracy of approximately 98.4% in the mouse Awake/Sleep classification task. Although this performance was slightly lower than that of the baseline mouse EEG model trained from scratch (approximately 99.8%)， the transfer learning model still maintained an accuracy exceeding 98%. These findings suggest that the time-frequency features learned from human EEG data do not become entirely ineffective due to species differences and retain a certain degree of cross-species transferability. The results further support the feasibility of utilizing human EEG pre-trained models as feature extractors for mouse EEG classification tasks. In summary， this study verified the effectiveness of combining deep learning techniques with EEG time-frequency representations for sleep classification tasks and preliminarily explored the feasibility of cross-species feature transfer. The findings demonstrate that convolutional neural networks can effectively extract EEG characteristics and possess potential applicability across different species， providing valuable insights for future EEG-based sleep analysis and neuroscience research. Overall， this study verifies the effectiveness of combining deep learning with EEG time-frequency representations for sleep classification tasks and preliminarily explores the feasibility of cross-species feature transfer. The findings suggest that convolutional neural networks can not only effectively extract discriminative EEG features but also possess potential for cross-species applications， providing valuable insights for future EEG sleep analysis and neuroscience research.

運用生物資訊學平台探討子宮內膜異位症之分子機制與其候選藥物篩選;Bioinformatics-Based Exploration of Molecular Mechanisms and Potential Candidate Drug Screening in Endometriosis

Tue, 20 Oct 2026 06:22:12 GMT

title: 運用生物資訊學平台探討子宮內膜異位症之分子機制與其候選藥物篩選;Bioinformatics-Based Exploration of Molecular Mechanisms and Potential Candidate Drug Screening in Endometriosis abstract: 子宮內膜異位症（endometriosis）為一種雌激素依賴性慢性發炎疾病，儘管其病理進程已被廣泛探討，但病灶局部的微環境重塑與免疫逃逸機制仍未完全闡明。為探究其核心致病機制，本研究整合三組 GEO 微陣列（microarray）資料集，鑑定出 231 個一致性的差異表達基因（differentially expressed genes， DEGs）。功能富集分析顯示，DEGs 高度集中於「細胞外微環境重塑」與「補體級聯活化」。獨立的CIBERSORTx 分析揭示，病灶呈現高度免疫抑制（immunosuppressive）狀態，特徵為活化型自然殺手細胞（natural killer cell， NK cell）顯著減少及 M2 型巨噬細胞（M2 macrophage）浸潤趨勢增加。結合蛋白質交互作用（protein-protein interactions， PPI）網路，本研究篩選出 10 個核心分泌蛋白並提出調控模型：病灶可能透過 FMOD 與 PRELP 介導補體驅動的免疫逃逸（誘導 NK 細胞失能並阻斷膜攻擊複合體 [membrane attack complex， MAC] 形成）；同時藉由 LTBP2、MFAP4 與 DCN 之間的交互作用調節 TGF-b 的生物可利用性，進而驅動組織纖維化與局部免疫逃逸。最後，L1000CDS2 藥物重定位預測出 HDAC 抑制劑（histone deacetylases inhibitors， HDAC inhibitors）（如 Trichostatin A）與蛋白質合成抑制劑（如具備阻斷囊泡運輸作用的 Brefeldin A）為潛在候選藥物，有望藉此干預調節病灶的「表觀遺傳失調」與「高分泌表型」。總結而言，本研究透過生物資訊分析，初步描繪出由細胞外分泌蛋白主導的微環境重塑機制，建立了「細胞外基質異常」與「局部免疫失能」之間的初步理論連結，並為未來非荷爾蒙標靶治療提供了客觀的數據參考依據。;Endometriosis is an estrogen-dependent chronic inflammatory disease; despite extensive pathological research， mechanisms underlying lesional microenvironment remodeling and immune evasion remain incompletely elucidated. To explore its core pathogenic network， this study integrated three GEO microarray datasets， identifying 231 consistent differentially expressed genes (DEGs). Functional enrichment showed these DEGs heavily concentrate in "extracellular microenvironment remodeling" and "complement cascade activation". Independent CIBERSORTx analysis revealed a highly immunosuppressive lesional state， characterized by significantly decreased activated natural killer (NK) cells and increased M2 macrophage infiltration. Integrating protein-protein interactions (PPI) networks， this study screened 10 core secreted proteins and proposed a regulatory model: lesions may mediate complement-driven immune evasion via FMOD and PRELP (inducing NK cell exhaustion and blocking membrane attack complex [MAC] formation); simultaneously， through LTBP2， MFAP4， and DCN interactions， lesions modulate TGF-b bioavailability to drive tissue fibrosis and local immune evasion. Finally， L1000CDS2 drug repurposing predicted histone deacetylases (HDAC) inhibitor (e.g.， Trichostatin A) and protein synthesis inhibitors (e.g.， Brefeldin A， a vesicle transport blocker) as potential candidates to modulate lesional "epigenetic dysregulation" and "hyper-secretory phenotype". In conclusion， this bioinformatics study delineates an extracellular secreted protein-driven microenvironment remodeling network， theoretically linking "extracellular matrix abnormalities" with "local immune dysfunction，" providing objective data for future non-hormonal targeted therapies.