鑒於資料科學過去二十年的蓬勃發展，生物醫學領域對於新興的資料處理概念與工具的興趣日益增加，然而生醫數據普遍面臨高維度小樣本（p >> n）、數據品質差異劇烈與專業知識壁壘等核心挑戰，現有的自動化機器學習（AutoML）工具或缺乏對數據預處理流程的深度整合、或受限於計算效率，難以滿足研究需求。本研究開發開源工具組 PineBioML，旨在為複雜生醫數據提供端到端的解決方案。 本研究的核心技術有三：（一）端到端的流程探索管理工具Pine，自動探索從預處理到模型的完整資料處理流水線，避免冗於運算的同時確保資料流嚴格隔離。（二）透過集成技巧結合五種特徵選擇方法，在處理高維數據時換取顯著的運算效率提升。（三）模型超參數優化的過擬合預防策略，在優化過程中修正樂觀偏差以提升泛化能力，增加機器學習模型於異質資料的適應性與穩健性。在此之上，PineBioML亦具備高度的程式介面兼容性，結合scikit-learn與pandas並且囊括SHAP、UMAP、PCA和PLS-DA等解釋工具，提供豐富的模型解釋與資料視覺化功能，完善作為一個便捷工具的職能。PineBioML開源於：https://github.com/ICMOL/PineBioML。 實際表現上，在 JADBio Benchmark 的 360 個建模任務的結果中，PineBioML 以0.883的平均測試 AUC分數優於 TPOT 0.872、FLAML 0.831以及PyCaret的0.616。在處理大型數據集時，建模速度領先最高達 40 倍。此外，在38個生醫研究數據集中，二分類任務的平均排名亦顯著優於MLJAR、FLAML 及 PyCaret。 本研究證實 PineBioML 具備作為生醫研究數據分析便捷工具的資質，有望加速研究者從複雜數據中提取關鍵科學資訊之進程。;The application of machine learning in biomedical research is frequently constrained by the high-dimensional small-sample (p≫n) regime， extreme data heterogeneity， and the substantial technical barriers involved in pipeline construction. Existing automated machine learning (AutoML) frameworks either lack sufficient integration with data preprocessing or exhibit prohibitive computational inefficiencies when processing high-dimensional datasets. This study presents PineBioML， an open-source， end-to-end AutoML toolkit specifically engineered to address these domain-specific challenges. PineBioML incorporates three primary technical modules. The Pine module serves as a pipeline management system that utilizes a depth-first search (DFS) strategy to automatically identify optimal processing-to-modeling pathways while maintaining rigorous data flow isolation to prevent leakage. The ensemble feature selection module aggregates five distinct embedding and filtering feature selection methods， enhancing feature selection stability and achieving up to a 40-fold modeling speedup in high-dimensional spaces. In addition， the bias-corrected optimization module integrates Optuna with the Tibshirani correction to mitigate the optimistic bias commonly observed in iterative hyperparameter searches， thereby improving model robustness across heterogeneous datasets. PineBioML is open-sourced at: https://github.com/ICMOL/PineBioML. Built upon the scikit-learn and pandas ecosystems， PineBioML provides high interface compatibility and advanced interpretability through SHAP， UMAP， and PLS-DA. Empirical evaluation on the JADBio Benchmark (360 tasks) demonstrates that PineBioML achieves a mean test AUC of 0.883， significantly outperforming TPOT (0.872)， FLAML (0.831)， and PyCaret (0.616). Furthermore， extensive testing on the BioML38 dataset confirms that the toolkit consistently achieves superior average rankings in both binary and multi-class classification tasks. These results validate PineBioML as an efficient and accessible platform for accelerating scientific discovery from complex biomedical data.
