實驗結果指出：靜態成熟度（GA、BBW、PMA）、生理訊號（HR、RR、SBP、DBP）、流體輸入/輸出量、imputation flags 與 surgery flag 共 8 個 feature groups（37 features）構成最小有效特徵集；完整 event-based 特徵（echo 即時、echo 歷史、medication）在當前 sparse-label 設定下未提供穩定邊際資訊，移除後 validation MAE 反而下降 2–6%。Reduced Optimal 模型於 hold-out test 之 5-seed ensemble 達 total MAE 1.382 mm、generalization gap −0.059 mm（由 Original Big 之 +0.429 mm 改善），通過 |Δ_gen| < 0.10 mm 之檢查；但其 Spearman ρ_S 由 0.36/0.25（T+1/T+4）降至 0.05/0.08，呈現 mean-predictor 退化。進一步的 Balanced A4 refinement 僅加回單一 last_echo_size，並以原始 A4 容量訓練，5-fold CV total MAE 為 1.113 ± 0.124，hold-out total MAE 降至 1.301 mm，T+1 Spearman 提升至 0.397；T+4 Spearman 仍僅 0.164，顯示遠期預測仍受標籤稀疏與事件時序限制。
本研究之主要貢獻為三點：(1) 將 PDA 預測自二元分類重新定義為連續尺寸回歸；(2) 將 GRU-D 之 mask-based 思想對偶延伸至標籤之時間稀疏並設計 masked MAE 損失；(3) 建立 LOGO + AOGI + noise-aware 三重特徵消融驗證框架，並透過 Balanced A4 展示容量與特徵精簡之折衷。本論文並未主張提出新模型架構或宣稱 SOTA 表現，而是於固定 LSTM 家族與真實小樣本醫療資料下，以可重現流程量化稀疏標籤條件下深度學習的能力邊界，並主張此類任務應同時報告 MAE、Spearman、generalization gap 與 calibration，以避免單一指標之誤判。;Patent ductus arteriosus (PDA) is one of the most common cardiovascular abnormalities in extremely low birth weight preterm infants. Clinically， PDA size (mm) measured by echocardiography (echo) is used as a continuous indicator of severity. However， echo is not a daily routine examination: most patient-days during a neonatal intensive care unit (NICU) stay carry no valid measurement， making "future PDA size prediction" a canonical regression task with temporally sparse labels. Existing AI work on PDA prediction has overwhelmingly framed the task as binary classification (hsPDA vs. non-hsPDA， treatment success vs. failure); this thesis intentionally avoids the subjectivity of classification thresholds and redefines the task as continuous size regression.
The study cohort comprises 212 preterm infants observed over a 30-day window (6，360 patient-days in total). The prediction targets are PDA size at T+1 and T+4 days， whose valid-label coverage is only 7.0% and 5.0% respectively. The methodology is organised into three layers. (i) Data layer: a reproducible day-level processing pipeline is built， covering SHA-256 deidentification， free-text event parsing， age-of-life alignment， echo deduplication， and 49-dimensional feature engineering organised into 11 clinically motivated feature groups. A binary mask and a masked MAE loss are introduced at the loss-function level， dualising the mask-based treatment of "informative missingness" in GRU-D from input features to temporally sparse labels. (ii) Model layer: the LSTM family is fixed as the backbone. Original Big (266k parameters) and Reduced Small (48k parameters) are first compared to diagnose overfitting， and a Balanced A4 refinement is then constructed by combining the higher-capacity A4 backbone with a compact feature set. (iii) Interpretability layer: a dual-view group-level ablation framework combining Leave-One-Group-Out (LOGO) and Add-One-Group-In (AOGI) is established， together with a selection rule for the minimum effective feature set and a 5-seed noise-aware sensitivity analysis.
Results show that static maturity (GA， BBW， PMA)， vital signs (HR， RR， SBP， DBP)， fluid input/output， imputation flags， and surgery flags — eight feature groups and 37 features in total — constitute the minimum effective feature set. Full event-based groups (echo-today， echo-history， medication) provide no stable marginal benefit in the current sparse-label setting; their removal even lowers validation MAE by 2–6%. The Reduced Optimal model achieves a 5-seed ensemble total MAE of 1.382 mm on the hold-out test set with a generalization gap of −0.059 mm (improved from +0.429 mm of Original Big)， satisfying the |Δ_gen| < 0.10 mm overfitting-resolution criterion. However， its Spearman ρ_S drops from 0.36/0.25 (T+1/T+4) to 0.05/0.08， exposing mean-predictor degradation. The subsequent Balanced A4 refinement adds back only last_echo_size and uses the original A4 capacity; it achieves a 5-fold CV total MAE of 1.113 ± 0.124， a hold-out total MAE of 1.301 mm， and improves T+1 Spearman to 0.397， although T+4 Spearman remains limited at 0.164.
The main contributions are threefold: (1) redefining PDA prediction from binary classification to continuous size regression; (2) extending the mask-based treatment of informative missingness from input features to temporally sparse labels via a masked MAE loss; and (3) establishing a triple ablation framework (LOGO + AOGI + noise-aware sensitivity) for clinical time-series LSTM models， with Balanced A4 illustrating a capacity-feature trade-off. This thesis does not claim a new architecture nor state-of-the-art performance; instead， under a fixed LSTM family and a realistic small-sample medical dataset， it offers a reproducible quantification of the capability boundary of deep learning under sparse supervision， and argues that such tasks should be reported with MAE， Spearman， generalization gap and calibration jointly， to avoid misjudgement based on any single metric.
