BOIN Safety Rules: “Cleared for Escalation” Is Not “Declared Safe”

BOIN 安全規則：升量許可 ≠ 劑量宣告安全

English

One of the most consequential misunderstandings in phase I oncology trial operations is a semantic one: the moment a cohort is cleared to escalate to the next dose level, the team often begins treating the current dose as having been declared safe. This conflation is understandable — if a dose was tolerated well enough to move on, it seems natural to use it as a foundation for what comes next. But in the world of adaptive Bayesian dose escalation, this reasoning contains a hidden flaw that can expose patients in combination studies, regional bridging trials, or intra-patient escalation to doses that the data do not yet fully support.

A 2026 simulation study in Pharmaceutical Statistics made this point formally for the Bayesian Optimal Interval (BOIN) design. The authors modeled the operating characteristics of BOIN trials under different escalation strategies and examined how often a dose that was “cleared for escalation” at some early point in the trial was later de-escalated — that is, found to be too toxic once more data accumulated. The finding was sobering: because BOIN is an adaptive design that allows repeated re-evaluation, a dose that cleared an early cohort assessment is not necessarily the same as a dose that the entire trial ultimately endorses. The more aggressively the team escalated from a cleared dose — for instance, by immediately launching a combination cohort or allowing prior patients to escalate within themselves — the more patients ended up having been exposed to a dose that was eventually reconsidered. The simulation recommended that before a dose can be treated as genuinely safe for secondary uses, there should be at least some evidence from the dose level above it, typically at least three patients evaluated at that higher level, to provide a clearer safety bracket.

The clinical lesson here is about institutional culture as much as statistics. In busy multi-center phase I trials, there is enormous pressure to move quickly: sponsor timelines, patient enrollment windows, and the practical difficulty of keeping a biostatistician engaged across months of accrual all push toward treating any positive safety signal as provisional confirmation. When a safety review committee convenes and the BOIN decision table says “escalate,” the natural instinct is to immediately deploy that cleared dose in every parallel activity the sponsor has planned. The statistician’s model, however, is making a more limited claim: that the observed DLT count at this dose falls within the acceptable interval, and that the best next step for learning purposes is to go higher. It is not certifying that this dose level is permanently safe for independent clinical use.

This distinction becomes critically important in three specific situations. The first is combination therapy development. If a drug has cleared dose level three in a single-agent trial and the team wants to launch a combination cohort starting at dose level three, they are effectively treating that dose as a safe monotherapy baseline. But the data supporting dose level three may rest on only six patients whose follow-up is still maturing. Adding a second drug on top of an inadequately characterized dose can amplify toxicity in ways the single-agent data cannot predict. The second situation is regional or bridging escalation, where a dose cleared in one geographic population is used as the starting point for a separate cohort in another region. Population PK differences, different patient selection, or varying treatment supportive care standards mean that “cleared in the US cohort” is not a universal safety certificate. The third situation is intra-patient dose escalation — allowing individual patients who tolerated a lower dose to move up to a higher one during the same trial. This is the subject of the IP-CRM design (intra-patient dose-escalation continual reassessment method), which explicitly models carry-over effects from earlier doses on subsequent toxicity. Without this kind of formal accounting, it is easy to attribute a toxicity at dose level four to the current dose when it may actually reflect cumulative exposure that began at dose level two.

The broader teaching is about how to read the vocabulary of phase I trial reports. The language of escalation decisions — “cohort cleared,” “dose approved,” “safety review passed,” “escalation permitted” — all describes operational decisions, not evidentiary milestones. An evidentiary milestone would say something like “dose level three has been observed in 18 patients with a median follow-up of 12 weeks, the upper confidence bound on DLT probability is 28%, and the observed exposure-response relationship is consistent with preclinical predictions.” The gap between operational language and evidentiary language is where patient safety decisions get made, and clinical investigators who sit on dose escalation safety committees need to be able to push past the former to demand the latter.

Practical guidance for journal club: when reading a phase I paper, identify three distinct claims about any given dose — (1) it was cleared for escalation, (2) it was selected as the RP2D, and (3) it was called safe enough to use in a combination or expansion setting. These are three very different evidentiary claims, and the paper should support each one separately. If the RP2D rationale simply says “tolerated by the majority of patients in the escalation cohort,” without citing the number of patients, DLT maturity, exposure data, or comparison to an alternative dose, the dose selection is not adequately supported regardless of how sophisticated the escalation design was. The statistical design is only as good as the evidence it generated, and that evidence must be read critically — not accepted as a guarantee because a model approved the escalation step.

中文

第一期腫瘤試驗操作中最具有後果的誤解之一，其實是語義上的誤解：當一個 cohort 被允許升到下一個劑量層級的那一刻，研究團隊往往開始把目前這個劑量視為「已宣告安全」。這種混淆是可以理解的——如果一個劑量的耐受性足以讓試驗繼續前進，把它當作後續工作的基礎似乎很自然。但在適應性貝氏劑量升階的世界裡，這個推論隱藏著一個陷阱，可能讓組合療法研究、區域性橋接試驗或病人內升量的受試者，暴露在資料尚未充分支持的劑量下。

2026 年 Pharmaceutical Statistics 的一篇模擬研究針對 BOIN 設計正式提出了這一點。作者對不同升階策略下的 BOIN 試驗操作特性進行建模，並檢驗在試驗早期被「允許升階」的劑量，後來又被降階（即隨著更多資料積累被判定為過毒）的頻率。結論令人省思：由於 BOIN 是允許反覆重新評估的適應性設計，在早期 cohort 審查中通過的劑量，不一定等同於整個試驗最終認可的劑量。團隊從一個已通過的劑量越積極升階——例如立即啟動組合 cohort，或讓舊病人在自身內部升量——最終就有越多病人曾暴露於一個後來被重新考慮的劑量。模擬結果建議：在一個劑量可被視為真正安全、可用於次要用途之前，應有來自上一個劑量層級的一些證據，通常至少要有三位病人在更高劑量層級完成評估，以提供更清晰的安全邊界。

這裡的臨床教訓不只關乎統計，也關乎機構文化。在繁忙的多中心第一期試驗中，壓力無所不在：贊助商的時程、病人入組窗口、讓統計師在數月收案過程中保持參與的實際困難，都推著大家把任何正面的安全訊號視為暫時性確認。當安全審查委員會召開，BOIN 決策表顯示「升階」，本能的反應是立即把那個已通過的劑量部署到贊助商計畫的所有並行活動中。然而統計師的模型做出的是一個更有限的聲明：在此劑量觀察到的 DLT 數落在可接受區間內，且從學習的角度來看，最好的下一步是往更高劑量走。這並不是在證明這個劑量層級永久地可以安全地獨立用於臨床。

這個區別在三個特定情境中變得極其重要。第一是組合療法開發：如果一個藥物在單藥試驗中通過了第三個劑量層級，而團隊想從第三劑量層級啟動組合 cohort，他們實際上是在把那個劑量當作安全的單藥基線。但支持第三劑量層級的資料可能只有六位追蹤期仍在成熟中的病人。在一個尚未充分描述的劑量上再加一個藥物，可能以單藥資料無法預測的方式放大毒性。第二是區域性或橋接升階：在一個地理族群中通過的劑量，被用作另一個地區 cohort 的起始點。族群藥物動力學差異、不同的病人選擇標準或不同的支持性照護標準，意味著「在美國 cohort 通過」不是普世的安全證書。第三是病人內劑量升階——允許在同一試驗中，對較低劑量耐受良好的個別病人升到較高劑量。IP-CRM（病人內升量連續再評估法）設計明確對來自早期劑量的遞延效應（carry-over effect）進行建模。沒有這種正式計算，很容易把第四劑量層級的毒性歸因於當前劑量，而它實際上可能反映的是從第二劑量層級開始累積的暴露。

更廣泛的教學是關於如何閱讀第一期試驗報告的語彙。升階決策的語言——「cohort 已清關」、「劑量已核准」、「安全審查通過」、「允許升階」——描述的都是操作性決策，不是證據里程碑。一個證據里程碑應該說：「第三劑量層級已在 18 位病人中觀察，中位追蹤時間 12 週，DLT 機率的上信賴區間為 28%，觀察到的暴露-反應關係與非臨床預測一致。」操作性語言與證據性語言之間的落差，正是病人安全決策被制定的地方。坐在劑量升階安全委員會的臨床研究者必須能夠突破前者，要求後者。

期刊討論實用指引：閱讀一篇第一期論文時，針對任何一個特定劑量識別三個不同的聲明——(1) 它被允許升階；(2) 它被選為 RP2D；(3) 它被認定足夠安全可用於組合或擴增設定。這是三個非常不同的證據聲明，論文應分別支持每一個。若 RP2D 的依據只說「多數患者在升階 cohort 中耐受」，而沒有引用病人數、DLT 成熟度、暴露資料或與替代劑量的比較，那麼無論升階設計有多複雜，劑量選擇都不算有充分支撐。統計設計的品質，只體現在它所產生的證據品質上，而那些證據必須被批判性地閱讀——不能因為一個模型批准了升階步驟，就把它當成保證。

Key Concepts | 核心概念

術語	定義
Cohort escalation decision	操作層級決策：依規則可進入下一劑量 cohort
Dose declared safe	證據層級聲明：該劑量有足夠安全邊界支持獨立使用
RP2D	Recommended Phase 2 Dose，建議第二期劑量
IP-CRM	Intra-patient CRM，納入病人內升量的連續再評估法
Carry-over effect	前一劑量對後續毒性的遞延影響
Operating characteristics	設計的統計表現特性
De-escalation	降階，試驗回到較低劑量

Related Pages | 相關頁面

• dose-escalation-design-comparison — 3+3、BOIN、CRM 設計的全面比較
• tite-boin-late-toxicity — 晚發毒性與 pending DLT 的處理
• backfill-dose-optimization — 回填策略與多劑量候選者的比較
• model-informed-dose-optimization — MIDD 與模型輔助劑量最佳化