TITE-BOIN: Managing Late-Onset Toxicity and Fast Accrual in Phase I Trials
TITE-BOIN：處理晚發毒性與快速收案的第一期試驗設計
English
Imagine a phase I trial for a novel ADC where the first cohort completes treatment and — by all early indications — tolerates the drug well. The safety review team convenes, the BOIN decision table says escalate, and three new patients are enrolled at the next dose level before cohort one has finished their full DLT observation window. Six weeks later, two patients from cohort one develop grade 3 interstitial lung disease. The new cohort is already receiving a higher dose. This scenario is not hypothetical — it is a real operational hazard for any drug whose toxicity unfolds over weeks or months rather than days, and for trials where accrual speed outpaces the biology of the drug’s toxic effects.
The conventional BOIN design assumes that DLT observation is complete before the next dose decision is made. In practice, this assumption breaks down whenever two conditions coincide: the drug has a delayed toxicity mechanism, and patients are enrolling faster than the DLT window can close. Immune-related toxicities from checkpoint inhibitors, cumulative hematologic suppression from ADCs, cytokine-mediated effects from T-cell engagers, and late-onset hepatic or pulmonary toxicities from targeted therapies can all emerge well outside a standard 28-day DLT window. For such drugs, the decision to escalate may genuinely need to be made before all the relevant toxicity information is in — not because the team is being careless, but because delaying accrual indefinitely to wait for all data would make the trial practically unrunnable.
Time-to-Event BOIN (TITE-BOIN) was developed to address exactly this situation. The original design, published in Clinical Cancer Research in 2018, allows the dose escalation decision to incorporate information from patients who are still within their DLT observation window by weighting their contribution according to how much of the window they have completed. A patient who has completed 80% of the DLT window and has not yet experienced toxicity contributes more information than one who is only 10% through. This is not just a statistical convenience — it reflects a genuinely more accurate description of the evidence available. A patient who has tolerated the drug for five weeks of a six-week DLT window has told us something real about tolerability, even if the sixth week has not yet arrived.
A 2025 Clinical Cancer Research paper from FDA and MD Anderson investigators then addressed the practical question of how to operationalize TITE-BOIN safely. Their key contribution was translating the statistical framework into concrete operational thresholds that safety committees can actually apply. The most important: if fewer than 51% of patients at the current dose level have completed DLT evaluation, the team should generally pause accrual rather than escalate. A second threshold addresses what happens when TITE-BOIN recommends escalation but many patients at the current dose have very limited follow-up time — specifically, if the shortest follow-up among pending patients is less than 25% of the DLT window, escalation should be reconsidered. These are not arbitrary numbers. They represent the point at which the pending data is so thin that the statistical weighting procedure can no longer provide meaningful protection against a delayed toxicity signal that has not yet manifested.
The backfilling companion to TITE-BOIN is BF-BOIN (Backfilling BOIN), described in a 2024 Clinical Cancer Research paper. BF-BOIN addresses a different but related problem: in trials where escalation happens efficiently, the lower and middle dose levels end up with very few patients — sometimes only three — making it nearly impossible to estimate whether those doses might offer a better benefit-risk profile than the highest tolerated dose. BF-BOIN allows additional patients to be enrolled at dose levels that have cleared the safety threshold and shown some early activity signal, while the main escalation sequence continues toward higher doses. The practical effect is a richer dataset across the dose range, which is precisely what Project Optimus requires when the final RP2D decision must consider not just toxicity but also efficacy, pharmacokinetics, tolerability during long-term treatment, and dose modification patterns.
For a clinician sitting on a safety review committee, the TITE-BOIN framework changes what questions to ask at each meeting. The relevant questions are no longer just “how many DLTs did we see?” but: what fraction of patients at the current dose have completed their full DLT observation window? Among those still pending, how long have they been followed? If delayed toxicity appears in the next two weeks from any pending patient, what would it mean for the escalation decision that was just made? Is there a plan in the protocol for exactly this scenario? These are questions about data maturity — and data maturity is the true currency of safe dose escalation in drugs with delayed toxicity mechanisms. A BOIN decision table is only as trustworthy as the completeness of the information it is processing.
The teaching synthesis from TITE-BOIN is not that late toxicity makes dose escalation impossible — it is that fast accrual and delayed toxicity together create a specific, recognizable risk that demands a pre-specified operational response. Trials that use TITE-BOIN should not simply report that they used the method; they should report what their pending data thresholds were, how many times the safety committee invoked those thresholds to pause or modify escalation, and what late-occurring toxicities were captured that would have been missed by conventional BOIN. That level of transparency is what differentiates a statistically sophisticated escalation design from one that is merely statistically named.
中文
想像一個針對新型 ADC（抗體藥物複合體）的第一期試驗：第一個 cohort 完成治療，根據所有早期指標，病人耐受良好。安全審查團隊召開會議，BOIN 決策表顯示升階，三位新病人已在第一 cohort 完成完整 DLT 觀察窗之前，就開始接受下一個劑量層級的治療。六週後，第一 cohort 有兩位病人出現第三級間質性肺炎。新 cohort 已在接受更高劑量。這個情境不是假設——它是任何毒性在數週或數月而非數天內展開的藥物，以及收案速度超過藥物毒性生物學的試驗，所面臨的真實操作風險。
傳統 BOIN 設計假設在做下一個劑量決策之前，DLT 觀察已經完成。在實務中，當兩個條件同時成立時，這個假設就會崩潰：藥物有延遲毒性機轉，且病人入組速度超過 DLT 觀察窗的關閉速度。免疫檢查點抑制劑的免疫相關毒性、ADC 的累積性骨髓抑制、T-cell engager 的細胞激素介導效應、以及標靶治療的晚發肝毒性或肺毒性，都可能在標準 28 天 DLT 觀察窗之後才出現。對這類藥物，升階決策可能確實需要在所有相關毒性資訊齊全之前做出——不是因為研究團隊不謹慎，而是無限期推遲收案等待所有資料，會讓試驗在實務上無法執行。
Time-to-Event BOIN（TITE-BOIN，時間事件貝氏最佳區間設計）正是為了解決這個問題而開發的。2018 年發表於 Clinical Cancer Research 的原始設計，允許劑量升階決策納入仍在 DLT 觀察窗內的病人資訊，方式是根據他們已完成觀察窗的比例加權其貢獻。一位已完成六週觀察窗的 80%（即五週）且尚未出現毒性的病人，貢獻的資訊多於一位只完成了 10% 的病人。這不只是統計上的便利——它反映了對現有證據更準確的描述。一位在六週 DLT 觀察窗的五週內耐受了藥物的病人，即使第六週尚未到來，已經告訴了我們一些關於耐受性的真實資訊。
2025 年一篇來自 FDA 和 MD Anderson 研究者的 Clinical Cancer Research 論文，進一步解決了如何安全地操作 TITE-BOIN 的實務問題。他們最重要的貢獻，是把統計框架轉化成安全委員會真正可以應用的具體操作門檻。最重要的一條：若目前劑量層級少於 51% 的病人已完成 DLT 評估，研究團隊通常應暫停收案而非升階。第二個門檻處理 TITE-BOIN 建議升階但當前劑量有許多病人追蹤時間非常有限的情況——具體而言，若待評估病人中最短追蹤時間少於 DLT 觀察窗的 25%，應考慮暫停升階。這些不是任意的數字，而是代表待觀察資料稀薄到統計加權程序無法再對尚未出現的延遲毒性訊號提供有意義保護的臨界點。
TITE-BOIN 的回填配套是 BF-BOIN（回填 BOIN），於 2024 年的 Clinical Cancer Research 論文中描述。BF-BOIN 解決的是一個不同但相關的問題：在升階進行順利的試驗中，較低和中間劑量層級最終病人數很少——有時只有三位——使得幾乎不可能估計這些劑量是否能提供比最高耐受劑量更好的效益-風險比例。BF-BOIN 允許在已通過安全門檻且有一些早期活性訊號的劑量層級額外入組病人，同時主要升階序列繼續向更高劑量前進。實際效果是整個劑量範圍有了更豐富的資料集，而這正是 Project Optimus 時代最終 RP2D 決策所需要的——不只看毒性，還要看療效、藥物動力學、長期治療的耐受性，以及劑量調整模式。
對於坐在安全審查委員會的臨床醫師而言，TITE-BOIN 框架改變了每次會議應提出的問題。相關問題不再只是「我們看到多少 DLT？」，而是：目前劑量有多大比例的病人已完成完整 DLT 觀察窗？那些仍在追蹤中的病人被追蹤了多久？若未來兩週任何待評估病人出現延遲毒性，這對剛做出的升階決策意味著什麼？protocol 是否對這個精確情境有預先規劃？這些是關於資料成熟度的問題——而資料成熟度是具有延遲毒性機轉藥物安全劑量升階的真正貨幣。BOIN 決策表的可信度，只有在它所處理的資訊足夠完整的前提下才成立。
TITE-BOIN 的教學總結不是說晚發毒性讓劑量升階變得不可能——而是說快速收案與延遲毒性共同創造了一個特定的、可識別的風險，這個風險需要預先指定的操作回應。使用 TITE-BOIN 的試驗不應只報告使用了這個方法；應該報告他們的待評估資料門檻是什麼、安全委員會援引這些門檻暫停或修改升階的次數，以及哪些晚發毒性被捕獲——若使用傳統 BOIN，這些毒性將會被遺漏。這種透明度，才是把一個統計上複雜的升階設計與一個只是統計上有名字的設計區分開來的標準。
Key Concepts | 核心概念
| 術語 | 定義 |
|---|---|
| TITE-BOIN | Time-to-Event BOIN，時間事件貝氏最佳區間設計，處理待評估資料的 BOIN 延伸版本 |
| BF-BOIN | Backfilling BOIN，回填貝氏最佳區間設計，允許在安全劑量層補充收案 |
| DLT window | Dose-Limiting Toxicity observation window，劑量限制毒性觀察窗 |
| Pending data | 仍在觀察窗內、尚未完成 DLT 評估的病人資料 |
| 51% 門檻 | 若少於 51% 病人完成 DLT 評估，建議暫停收案 |
| 25% 門檻 | 若待評估病人最短追蹤時間 <25% DLT 窗口，應重新考慮升階 |
| Late-onset toxicity | 晚發毒性，在標準觀察窗後才出現的毒性 |
Related Pages | 相關頁面
- dose-escalation-design-comparison — 3+3、BOIN、CRM、CFO 的全面比較
- boin-safety-rules-dose-declarations — 升量許可 vs 劑量宣告安全
- backfill-dose-optimization — 回填策略的完整邏輯
- model-informed-dose-optimization — MIDD 與 seamless phase I/II 設計