# Adv_Biostat (2020.12.25)

###### tags: `BioStat`

**intuition:** 從授課老師背景推估這門課會學到什麼？

# 石瑜
## 學歷
淡大統計系→當兵→Michigan碩士花2年學**Probability**→轉學到安娜堡分校讀**Bio**statistics博士4.5年→Vanderbilt大學

## 身分
- Vanderbilt Center for Quantitative Sciences
- 癌症中心
- FDA External Advisory Voting Member
- JAMA Oncology
- 挺新的子刊，卻有20幾的IF
- 才四個人在運行

## Vanderbilt大學 (南方小哈佛)
- 美國前10名的私立醫學院
- 美國前20名的學校 (U.S. News & World Report)
- 約5000的undergraduate
- 全校一萬人，醫院卻有三萬個教職員

---

成大統計系的[蘇佩芳](https://researchoutput.ncku.edu.tw/zh/persons/pei-fang-su)老師也在做生物計量，可諮詢

## Syllabus: Design & Data Analysis

- Design
- how to define verb (primary objective)
- how to balance strength and weakness (cost)
- how to decide sample size
- randomisation
- lab
- Markov chain Monte Carlo
- R applications in linear regression
- Model fitting
- Multiple imputation for missing values
- Variable selection

- Data Analysis
- three variables
- continuous, categorical, time-to-event
- error avoiding
- towards high dimensional data

[https://github.com/vubiostat/ps](https://github.com/vubiostat/ps)

---

## FDA case study

### FDA Anti-Infective Drug Advisory Committee Meeting - Dec 10, 2009

- performed by Center for Drug Evaluation and Research (CDER)
- 在Washington, D.C./Gaithersburg的hotel
- 現在已經搬到FDA自己的建築開會
- 有錄音錄影、記者席、書記
- 每次開會前，至少要花一周時間review paper
- 報告時只會講一種drug，但是藥廠有好幾個trial需要review
- 流程
- Applicant的Biometrics專家→FDA的Biometrics專家→Public Hearing→投票(同意/反對/棄權)
- transparent: 送審的藥廠、病人與家屬等利益相關者
- 壓力很大
- 一位家屬說到自己的女兒本來應該出席感謝藥物的通過，卻不幸在前一周去世了，哭光了整包面紙
- 另一位則提到加拿大似乎很可能通過該藥，財務負擔已經很重了，莫非還要冒著長途車程去加拿大買藥？
- 8:00 a.m. - 4:00 p.m.

[AGENDA](Adv_Biostat%20(2020%2012%2025)%20466c3587ee414cb8a1120f0ba39bde93/AGENDA%202cf0f1e14773469d897e810100075f27.csv)

---

## AZLI

- [trial](https://clinicaltrials.gov/ct2/show/NCT00104520): NDA 050-814, inhaled aztreonam, manufactured by Gilead Sciences

**VOTE:** Has the Applicant provided substantial evidence of the efficacy and safety of 75 mg **three** times daily of AZLI for the requested indication of improvement of respiratory symptoms and pulmonary function in cystic fibrosis patients with Pseudomonas aeruginosa? In your response, discuss the rationale for your answer.

### Background

CF is an autosomal recessive disease that affects approximately 30,000 people in the US.

- rare disease

Currently, tobramycin inhalation solution (TIS, TOBI)is the only FDA-approved inhaled antibiotic for use in CF. TIS was approved in the US in 1997 and has been widely prescribed by the CF community for the past decade.

- IV跟inhalation
- 之所以不用inhalation，是因為患者年紀小

However, TIS cannot be used in all patients with CF, and only ~40% use it chronically.

- 不適用於所有人

New classes of aerosolized antibiotics for the treatment of chronic Pseudomonas aeruginosa (PA) infection are needed.

On September 16, 2008, the Division of Anti-Infective and Ophthalmology Products (DAIOP) issued a Complete Response Letter to Gilead indicating that the clinical evidence was insufficient to establish efficacy of AZLI due to a pronounced regimen effect observed in the placebo twice daily (BID) and TID arms of Study 005.

- 這個藥已經被拒絕一次了，藥廠再上訴

The DAIOP also questioned the acceptability of the Cystic Fibrosis Questionnaire-Revised (CFQ-R) patient reported outcome (PRO) tool as the primary endpoint in Study 007 and stated that there was sparse evidence to support the proposed 75 mg dose and TID regimen.

Gilead subsequently submitted a second-level appeal to the Office of New Drugs on March 13, 2009.

The Office of New Drugs recommended that Gilead submit the new analyses relevant to the AZLI NDA in response to the September 16, 2008 Complete Response Letter. In addition, the Office of New Drugs recommended that the DAIOP present the application to the Anti-Infective Drugs Advisory Committee.

- 不像歐盟和加拿大，FDA沒有conditional approval，只有同意、反對、emergency use
- 上訴成功的契機是：歐盟通過了該藥的conditional approval
- 有謠傳在這場會議後，加拿大似乎也要開會了

---

%20466c3587ee414cb8a1120f0ba39bde93/Untitled.png)

Figure1. Phase 3 Programme Overview

- TIS: only FDA-approved inhaled antibiotic for use in CF

- 為什麼不是AZLI versus TIS?
- 又為什麼要先給TIS?

- AZLI/Placebo 分四組(randomisation ratio= 2:2:1:1)
- AZLI 75 mg TID (=3 times/daily)
- AZLI 75 mg BID (=3 times/daily)
- Placebo TID
- Placebo BID
- Non Hypothesis: no differences between AZLI and Placebo
- Primary Endpoint: 又感染的時間點
- data analysis中的三種variables之一：time-to-event (survival)
- 也可以是time to death、recurrence、hospitalisation

---

%20466c3587ee414cb8a1120f0ba39bde93/Untitled%201.png)

Figure 7. Study 005 Primary Endpoint: Time to Need with Regimen and Treatment Group (ITT)

上圖是網路上找到的示意圖（原圖有版權）

- 先搞清楚x-y軸：依Days-Survival Function作圖
- 0 Days時，100%的人是event-free
- event指的是inhaled or IV antibiotics
- clear definition needed
- 這裡被定義為給予TIS後
- 保證大家都是infection-free
- Arm分四組：Placebo BID、Placebo TID、AZLI BID、AZLI TID
- survival function=0.5時
- 有一半的受試者在..天時需要event
- 希望新藥的median time to need能postpone
- Median time to need: n.e.、54 days、n.e.、87 days
- p value
- BID: AZLI vs Placebo p=0.4269
- TID: AZLI vs Placebo p=0.0043

---

### Placebo difference (study 005)

The large observed difference between the placebo BID and TID arms was unexpected and warrants further investigation.
At baseline, patients on placebo TID were potentially at higher risk upon entry into the study compared to patients on placebo BID.
Adjusting for these baseline imbalances in a Cox model did not remove the statistical significance (p = 0.01 after adjustment) of the difference between the placebo arms.
Therefore, this baseline imbalance did not explain the observed difference, and consequently, the difference between the placebo groups was due either to unmeasured covariates or to chance.
Since key prognostic variables were measured and adjusted for, it is most likely that the difference was due to chance alone.

- Something Wrong
- BID: AZLI vs Placebo沒有差別
- Placebo: BID vs TID差別很大
- 歸因：嚴重病人剛好被assign到TID，健康病人剛好被assign到BID (randomisation沒做好)
- N太小，不敢做stratification
- appropriate method: adaptive minimization randomization
- explanation: 將線倆倆合併，變成AZLI vs Placebo的p value = 0.0070
- 合併是被允許的嗎? FDA投票的問題是要TID
- Regimen effect for primary endpoint?
- Gilead contends that the regimen effect should not invalidate the persuasive results of the primary hypothesis tested in this study and that the pooling of the placebo groups for analyses provides interpretable results.
- The pre-specified primary efficacy endpoint for Study 005 was a comparison of time for IV or inhaled anti-pseudomonal antibiotics between pooled AI (BID and TID) versus pooled placebo (BID and TID), based on the log rank test.
- The primary efficacy endpoint for Study 005 was met, and therefore secondary endpoints using the pre-specified comparison of pooled AZLI versus pooled placebo were explored.
- FDA ground:
- p-value: AZLI/Placebo的TID比Placebo的TID/BID還大(0.0043和0.0012)，那幹嘛吃藥？
- Sensitivity analyses were consistent with primary analysis
- Highly significant & influential effect from BID vs. TID regimen
- AZLI TID worse than Placebo BID (in event rates and time to event)
- No clear evidence of AZLI TID benefit over placebo
- Study results not reliable due to high overall dropout rates and the depedent nature of the dropouts
- More AZLI vs. Placebo (TSI) + more TID vs. BID droupouts
- Dropouts often due to AEs indicating dropouts may be sicker
- AZLI TID excluded more (sicker) dropouts, may benefit more
- Study results fail to show an AZLI TID benefit esp. when accounting for regimen effects.
Regimen effects were:
- Consistently stronger than the AZLI drug effect
- Detrimental to TID regiments (e.g. AZLI TID)
- 2° analyses failed to show an AZLI TID benefit. CFQ-R RSS & FEV1 had only marginal effects at Day 28. Also:
- No evidence of a sustained AZLI TID effect after Day 28
- AZLI TID was worse than placebo by Day 84
- Biases in time to need could not be clearly assessed using the Applicant's new post-hoc analyses
- S1 not informative, FEV1 slope treatment related
- S2 problematic given differences in time to need & FEV1. S2 also did not affect primary analysis findings.
- AZLI TID (those with events) had no FEV1 benefit at Day 42+ and had significantly greater hospitalization.
- Study 005 provided no evidence of an AZLI TID benefit.
Suggested possible harm from use of a TID regimen.
Findings unlikely due to chance:
- Prob.(detrimental TID regeimne effect in 1° analysis) = .0006
- Prob.(TID vs. BID dropouts in those censored early) = .0353

你的決定？會通過還是否決？

---

## Current use of RWD for Evidence Generation

- RWD (Real World Data)
- not well-controlled (or randomised)
- 對統計分析的嚴謹度要求更高
- 可以是EMRs (Electronic medical records)、EMR，或是來自病人身上的devices (iPhone測量的心跳血壓)、問卷...
- 前言
- 肺炎疫情的肆虐下，來不及進行臨床試驗，於是想要老藥新用
- 但是，如何評估效用？

### Studies of Off-label Drugs

- 相關文獻
- Clinical impact of COVID-19 on patients with cancer (CCC19): a cohort study
- on The Lancet
- Utilization of COVID-19 Treatments and Clinical Outcomes among Patients with Cancer: A COVID-19 and Cancer Consortium (CCC19) Cohort Study
- on Cancer Discovery
- Tocilizumab Treatment for Cytokine Release Syndrome in Hospitalized Patients With Coronavirus Disease 2019: Survival and Clinical Outcomes
- on CHEST Journal
- Scientific Rigor in the Age of COVID-19
- on JAMA Oncology

---

### Design Aspect

- Cancer Discovery
- CCC19: 癌症和COVID-19的共病

%20466c3587ee414cb8a1120f0ba39bde93/Untitled%202.png)

UpSet plot of treatment exposures.

- drugs
- Tocilizumab: 參見石瑜在CHEST的發表
- Remdesivir: 西雅圖的重症患者服用的隔天就退燒，因而引起關注，後續的試驗結果並不特別
- Hydroxychloroquine: 瘧疾藥
- Primary Objective: Hydroxychloroquine
- 如何設計實驗，例如w/ or w/o Hydroxychloroquine?

---

%20466c3587ee414cb8a1120f0ba39bde93/Untitled%203.png)

Table 2. Evaluation of 30-day all-cause mortality associated with hydroxychloroquine exposure, as compared with positive (treated) and negative (untreated) controls using different methodological approaches

- negative control是沒吃任何藥，藉以區分輕重症患者

---

- Power Analysis vs. Precision Analysis
- power analysis: for sample size decision in RCT
- precision analysis: for effect size estimation
- effect size受treatment difference影響

- Note that p-value can not tell you the clinically significant level!
- 𝒇𝒏 (𝒔𝒂𝒎𝒑𝒍𝒆 𝒔𝒊𝒛𝒆, 𝒗𝒂𝒓𝒊𝒂𝒕𝒊𝒐𝒏, 𝒄𝒍𝒊𝒏𝒊𝒄𝒂𝒍𝒊𝒏𝒂𝒍 𝒆𝒇𝒇𝒆𝒄𝒕 𝒔𝒊𝒛𝒆)

在RWD中永遠有bias，而randomisation是避免bias很好的手段。

---

### Data Analysis Aspect

Key Analytical Steps

- Quality control - NLP
- Missing data analysis

%20466c3587ee414cb8a1120f0ba39bde93/Untitled%204.png)

How to impute the missing SCL for patient # 5?

- techniques: multiple imputation
1. 拿age, sex, edu., race做regression，得到四個residual值
2. 從residual distribution抓一個component出來，加在最後一個的SCL
3. 即便其他variable都相同，猜測的SCL值還是有variation，減少bias
- Propensity score matching
- 用score代表病人的overall severity
- 確保baseline的balance
- Model building & evaluation

- **DON’T DO** multiple **uni**variate analyses to determine some potential.
- **DON’T DO** Stepwise variable selection.

- causal inference
- 只知道A和B是correlated已經不夠
- Sensitivity analysis
- Causal/Mediation analysis

---

### Conclusion

- Ignorance of scientific rigour can cause detrimental effects
- Everything we eat both causes and prevents cancer
- Avoid data dredging or data fishing