AI 英文潤稿改了什麼？實測對照

測試資料：實測紀錄載明執行日期為 2026 年 10 月 28 日，使用 Claude Sonnet 5.5。三段內容均為虛構。本文沿用既有實測輸出，未重新潤稿。實測紀錄未附完整逐字提示詞，本節呈現可核對的規則檔內容。

對照標記：改寫前的灰色刪除線，標示被刪除或替換的文字；改寫後的藍色粗體，標示新增或替換的文字。英文中的字句與數值不因標色而更動。整句重寫時，標記以逐字比對為準，前後稿共有的字可能不標色，請以整句閱讀。

In recent years, remote blood pressure monitoring has attracted increasing attention in primary care. In this study, we conduct extensive experiments on a wide range of patients to delve into the intricate effects of home telemonitoring. We enrolled 412 adults from 6 clinics and randomized them 1:1. Moreover, mean systolic blood pressure fell by 8.2 mmHg (95% CI 5.9 to 10.5) in the telemonitoring arm versus 3.1 mmHg in usual care (p = 0.003). Furthermore, these results demonstrate that telemonitoring significantly outperforms usual care, paving the way for a transformative paradigm in hypertension management. It is worth noting that 87% of participants completed follow-up. To the best of our knowledge, this is the first study to comprehensively examine this crucial problem.

We tested whether home blood pressure telemonitoring lowers systolic blood pressure more than usual care in primary care. We enrolled 412 adults from 6 clinics and randomized them 1:1. Mean systolic blood pressure fell by 8.2 mmHg (95% CI 5.9 to 10.5) in the telemonitoring arm and by 3.1 mmHg in usual care (p = 0.003). Of the participants, 87% completed follow-up. These results come from a single trial in 6 clinics.

Prior work reports that sleep restriction impairs working memory (Walker and Stickgold, 2006; [12]), but the effect sizes vary widely across paradigms [14, 15, 19]. In our sample (n = 58), 2-back accuracy was lower after restricted sleep (M = 0.78, SD = 0.09) than after normal sleep (M = 0.84, SD = 0.07), a difference of 0.06 (Cohen's d = 0.74, p = 0.012). Importantly, the data suggest that this deficit may be partly mediated by reduced vigilance — a pattern that is consistent with the lapse hypothesis, although the indirect effect (ab = 0.021, 95% CI [0.004, 0.046]) was small. Reaction times were log-transformed, y_i = β0 + β1 x_i + ε_i, and analyzed with a mixed model. Moreover, it is worth noting that the model converged in all 58 participants. Furthermore, this groundbreaking finding underscores the pivotal role of sleep in cognition, and our extensive analyses comprehensively confirm that sleep restriction universally impairs memory.

Prior work reports that sleep restriction impairs working memory (Walker and Stickgold, 2006; [12]), but the effect sizes vary widely across paradigms [14, 15, 19]. In our sample (n = 58), 2-back accuracy was lower after restricted sleep (M = 0.78, SD = 0.09) than after normal sleep (M = 0.84, SD = 0.07), a difference of 0.06 (Cohen's d = 0.74, p = 0.012). The data suggest that this deficit may be partly mediated by reduced vigilance, a pattern consistent with the lapse hypothesis. The indirect effect was small (ab = 0.021, 95% CI [0.004, 0.046]). Reaction times were log-transformed, y_i = β0 + β1 x_i + ε_i, and analyzed with a mixed model. The model converged in all 58 participants. These results are consistent with an effect of sleep restriction on working memory in this sample; they do not establish that sleep restriction impairs memory in general.

We propose a novel method for imputing missing values. Our method is more robust and achieves a large improvement over existing approaches. Samples were normalized to total protein. The results suggest that the method may be useful for proteomic data, and we hypothesize that the gain stems from the explicit noise model. Our approach is also fast.

We propose a method for imputing missing values. Our method is more robust than existing approaches and improves on them. Samples were normalized to total protein. The results suggest that the method may be useful for proteomic data, and we hypothesize that the gain stems from the explicit noise model. The method is also fast.

We prove that our method significantly outperforms all prior approaches.

Our method improves held-out accuracy by 4--7 points over the strongest prior approach (Table 3); the gain is significant at p < 0.01 by a paired test.