Clinical Trial Dataset Research Articles

Growth differentiation factor 15 (GDF15) predicts relapse free and overall survival in unresected locally advanced non-small cell lung cancer treated with chemoradiotherapy

IntroductionGrowth differentiation factor 15 (GDF15) is a cytokine of the TGFβ family. Here, we analyzed GDF15 levels in patients with locally advanced non-small cell lung cancer (LA-NSCLC) who participated in OCOG-ALMERA (NCT02115464), a phase II randomized clinical trial, that investigated metformin in combination with standard of care concurrent chemoradiotherapy (cCRT). OCOG-ALMERA was not able to demonstrate benefit in the metformin arm. Therefore, biomarker studies are needed to better define stratification parameters for future trials.MethodsPatients were randomized to treatment with platinum-based chemotherapy and concurrent chest radiotherapy (60–66 Gy), with or without metformin (2000 mg/d). The trial collected tumor volume parameters, survival outcomes, and patient blood plasma at baseline, during (weeks 1 and 6) and 6 months after cCRT. Plasma GDF15 levels were assayed with the ELISA method. Statistical analyses explored associations between GDF15, survival outcomes, and radiotherapy tumor volumes.ResultsBaseline plasma levels of GDF15 were elevated in study patients, they increased during cCRT (p < 0.001), and the addition of metformin was associated with a further increase (week 6, p = 0.033). Baseline GDF15 levels correlated with the radiotherapy gross target volume (GTV, p < 0.01), while week 1 of radiotherapy levels correlated with radiotherapy planned target volume (PTV, p < 0.006). In multivariate analysis, baseline plasma GDF15 was prognostic for poor relapse-free (RFS) and overall survival (OS) (p = 0.005 and p = 0.002, respectively).ConclusionsGDF15 is a plasma marker that responds to the treatment of unresected LA-NSCLC with cCRT and metformin. GDF15 levels correspond with tumor volume and increased GDF15 levels predict for poor RFS and OS. These results require validation in larger clinical trial datasets.

A Comprehensive Guide to Generating SDTM Demographic Datasets Using R in Clinical Trials

R is a programming language widely used for statistical analysis and data visualization, offering a flexible and interactive environment supported by various packages for data cleaning, tidying, and analysis. It is particularly relevant for professionals in mathematics and statistics, including biostatisticians and programmers in the pharmaceutical and biotech industries. R provides a robust array of user-developed packages that can efficiently manipulate complex datasets, such as those based on the Study Data Tabulation Model (SDTM). The popularity of R in data-related fields has surged exponentially over the past decade due to its open-source nature, powerful statistical capabilities, and advanced visualization tools. In this paper, we demonstrate a step-by-step approach to generating an SDTM Demographic (DM) dataset using R. The process leverages R packages such as sas7bdat, tidyverse, haven, parsedate, dplyr, tidyr, and Hmisc. We also provide a detailed procedure for setting up the R environment required for this process. While R has been extensively used for exploratory analysis in the pharmaceutical and biotech industries, its application in creating and analyzing clinical trial datasets, such as SDTM, has been limited. Traditionally, SAS® has been the preferred tool for generating clinical trial datasets. This paper explores R’s potential as a viable alternative, offering enhanced flexibility and cost- effectiveness in clinical trial.

7425 Biochemical Response, Symptom Control, and Safety of Oral Octreotide Capsules Treatment in Acromegaly: A Post Hoc Analysis of a Large Pooled Database from Three Phase 3 Clinical Trials

Abstract Disclosure: M. Fleseriu: Consulting Fee; Self; Novo Nordisk, Amryt Pharma Holdings Ltd, Ascend Therapeutics (A Besins Healthcare company). Speaker; Self; Xeris Pharmaceuticals, Inc., RECORDATI_RARE_DISEASES_INC., Ipsen Pharma SAS. S.L. Samson: Advisory Board Member; Self; Amryt. Research Investigator; Self; Amryt. N. Biermasz: Advisory Board Member; Self; Recordati, Pfizer Global R&amp;D. Grant Recipient; Self; Macro Registry Grant, Amryt. C.J. Strasburger: Advisory Board Member; Self; Sandoz, Ipsen, Recordati. Consulting Fee; Self; Ascendis, Amryt, Novo Nordisk, Merck, Sandoz, Recordati. Speaker; Self; Ipsen, Novo Nordisk, HRA-Pharma. S. Fuchs Orenbach: Employee; Self; Chiesi Global Rare Diseases. J.A. Crompton: Employee; Self; Chiesi Global Rare Diseases. S. Melmed: Advisory Board Member; Self; Ionis, Crinetics. Consulting Fee; Self; Ipsen. Grant Recipient; Self; Pfizer. Background: Treatment with oral octreotide capsules (OOC) was evaluated in three phase 3 trials (CH-ACM-01, OPTIMAL, MPOWERED) in patients with acromegaly who previously tolerated and responded to the injectable somatostatin receptor ligands (iSRLs) octreotide and lanreotide. Results demonstrated efficacy and safety despite differences in trial design. Pooling data from all three trials can provide further insight into the biochemical response, symptom improvement, and safety profile of OOC. Objective: Analyze the biochemical changes, symptom control, and safety of OOC through a pooled analysis of phase 3 clinical trial datasets. Methods: Patients included were aged ≥18 years with acromegaly and responded biochemically to iSRL therapy before the initiation of OOC (defined as IGF-I &amp;lt;1.3 × ULN and mean integrated GH &amp;lt;2.5 ng/mL for CH-ACM-01 and MPOWERED, and IGF-I ≤1.0 × ULN for OPTIMAL). Data analyzed were from the following periods: baseline until the end of the double-blind placebo-controlled phase of OPTIMAL (36 weeks); the run-in phase of MPOWERED (26 weeks); end of dose escalation (2-5 months; efficacy data) and end of fixed dose phase (up to 7 months, safety data) for CH-ACM-01. Effect of prior iSRL dose was analyzed using a random effects meta-analysis. Symptom severity was scored using the Acromegaly Index of Severity (AIS) tool. Results: Efficacy, symptom, and safety data were available for 318, 194, and 329 patients, respectively. The overall response rate was 65% with no effect of prior iSRL dose (p&amp;gt;0.05). Mean (SD) IGF-I change from baseline was +0.17 (0.32; n=214), +0.19 (0.39; n=81), and +0.22 (0.57; n=23) for patients with baseline IGF-I levels ≤1 × ULN, 1 to &amp;lt;1.3 × ULN, and ≥1.3 × ULN. Twenty percent (16/81) and 35% (8/23) of patients with baseline IGF-I of 1 to &amp;lt;1.3 x ULN and ≥1.3 x ULN shifted to an improved category of response. There was a median decrease of one symptom per patient (p&amp;lt;0.05) and a mean decrease of 0.8 in AIS (p&amp;lt;0.05). The proportion of OOC responders reporting ≤1, 2, or 3 moderate/severe acromegaly symptoms at baseline and month 6 decreased from 56% to 39%, 32% to 20%, and 20% to 12%. At month 6, improvement was seen for all individual symptoms (joint pain, swelling, sweating, fatigue, headache) of moderate/severe grade and was significant for joint pain (36% to 22%) and swelling (15% to 10%) (p&amp;lt;0.02). Most adverse events (AEs) were gastrointestinal (GI) in nature (54%, n=179/329) with a median onset and duration of 15 and 12 days for first reported GI events. Conclusion: Pooled analysis confirms overall maintenance of biochemical control in patients previously responding to iSRL therapy, with decreased symptom burden in OOC responders. OOC was generally well tolerated with no new safety signals. GI-related AEs were largely transient. These results support OOC as a therapeutic option for treating acromegaly. Presentation: 6/3/2024

Intranasal Versus Oral Treatments for Allergic Rhinitis: A Systematic Review With Meta-Analysis

BackgroundTreatments for allergic rhinitis include intranasal or oral medications. ObjectiveTo perform a systematic review with meta-analysis comparing the effectiveness of intranasal corticosteroids or antihistamines versus oral antihistamines or leukotriene receptor antagonists in improving allergic rhinitis symptoms and quality of life. MethodsWe searched four bibliographic databases and three clinical trial datasets for randomised controlled trials (i) assessing patients ≥12 years old with seasonal or perennial allergic rhinitis, and (ii) comparing intranasal corticosteroids or antihistamines versus oral antihistamines or leukotriene receptor antagonists. We performed a meta-analysis of the Total Nasal Symptom Score (TNSS), Total Ocular Symptom Score (TOSS), Rhinoconjunctivitis Quality of Life Questionnaire (RQLQ), development of adverse events, and withdrawals due to adverse events. Certainty of evidence was assessed using GRADE. ResultsWe included 35 studies, most of which assessed patients with seasonal allergic rhinitis and displayed an unclear risk of bias. Superiority of intranasal treatments was found for all assessed outcomes. Intranasal corticosteroids were more effective than oral antihistamines at improving the TNSS (MD=-0.86; 95%CI=-1.21;-0.51; I2=70%), TOSS (MD=-0.36; 95%CI=-0.56;-0.17; I2=0%) and RQLQ (MD=-0.88; 95%CI=-1.15;-0.61; I2=0%), being mostly associated with clinically meaningful improvements. Superiority of intranasal corticosteroids at improving the TNSS was also found against oral leukotriene receptor antagonists (MD=-1.05; 95%CI=-1.33;-0.77). Intranasal antihistamines were more effective than oral antihistamines at improving the TNSS (MD=-0.47; 95%CI=-0.81;-0.14; I2=0%) and RQLQ (MD=-0.31; 95%CI=-0.56;-0.06; I2=0%). ConclusionsRandomized controlled trials suggest that intranasal treatments are more effective than oral treatments at improving symptoms and quality of life in seasonal allergic rhinitis.

Using multistate models with clinical trial data for a deeper understanding of complex disease processes.

A clinical trial represents a large commitment from all individuals involved and a huge financial obligation given its high cost; therefore, it is wise to make the most of all collected data by learning as much as possible. A multistate model is a generalized framework to describe longitudinal events; multistate hazards models can treat multiple intermediate/final clinical endpoints as outcomes and estimate the impact of covariates simultaneously. Proportional hazards models are fitted (one per transition), which can be used to calculate the absolute risks, that is, the probability of being in a state at a given time, the expected number of visits to a state, and the expected amount of time spent in a state. Three publicly available clinical trial datasets, colon, myeloid, and rhDNase, in the survival package in R were used to showcase the utility of multistate hazards models. In the colon dataset, a very well-known and well-used dataset, we found that the levamisole+fluorouracil treatment extended time in the recurrence-free state more than it extended overall survival, which resulted in less time in the recurrence state, an example of the classic "compression of morbidity." In the myeloid dataset, we found that complete response (CR) is durable, patients who received treatment B have longer sojourn time in CR than patients who received treatment A, while the mutation status does not impact the transition rate to CR but is highly influential on the sojourn time in CR. We also found that more patients in treatment A received transplants without CR, and more patients in treatment B received transplants after CR. In addition, the mutation status is highly influential on the CR to transplant transition rate. The observations that we made on these three datasets would not be possible without multistate models. We want to encourage readers to spend more time to look deeper into clinical trial data. It has a lot more to offer than a simple yes/no answer if only we, the statisticians, are willing to look for it.

Disability and Recurrent Stroke Among Participants in Stroke Prevention Trials

Stroke secondary prevention trials have disproportionately enrolled participants with mild or no disability. The impact of this bias remains unclear. To investigate the association between poststroke disability and the rate of recurrent stroke during long-term follow up. This cohort study is a post hoc analysis of the Prevention Regimen For Effectively Avoiding Second Strokes (PRoFESS) and Insulin Resistance Intervention After Stroke (IRIS) secondary prevention clinical trial datasets. PRoFESS enrolled patients from 2003 to 2008, and IRIS enrolled patients from 2005 to 2015. Data were analyzed from September 23, 2023, to May 16, 2024. The exposure was poststroke functional status at study baseline, defined as modified Rankin Scale (mRS; range, 0-5; higher score indicates more disability) score of 0 vs 1 to 2 vs 3 or greater. The primary outcome was recurrent stroke. The secondary outcome was major cardiovascular events (MACE), defined as recurrent stroke, myocardial infarction, new or worsening heart failure, or vascular death. A total of 20 183 PRoFESS participants (mean [SD] age, 66.1 [8.5] years; 12 931 [64.1%] male) and 3265 IRIS participants (mean [SD] age, 62.7 [10.6] years; 2151 [65.9%] male) were included. The median (IQR) follow-up was 2.4 (1.9-3.0) years in PRoFESS and 4.7 (3.2-5.0) years in IRIS. In PRoFESS, the recurrent stroke rate was 7.2%, among patients with an mRS of 0, 8.7% among patients with an mRS of 1 or 2, and 10.6% among patients with an mRS of 3 or greater (χ22 = 27.1; P < .001); in IRIS the recurrent stroke rate was 6.4% among patients with an mRS of 0, 9.0% among patients with an mRS of 1 or 2, and 11.7% among patients with an mRS of 3 or greater (χ22 = 11.1; P < .001). The MACE rate was 10.1% among patients with an mRS of 0, 12.2% among patients with an mRS of 1 or 2, and 17.2% among patients with an mRS of 3 or greater (χ22 = 103.4; P < .001) in PRoFESS and 10.9% among patients with an mRS of 0, 13.3% among patients with an mRS of 1 or 2, and 15.3% among patients with an mRS of 3 or greater (χ22 = 5.8; P = .06) in IRIS. Compared with patients with an mRS of 0, patients with an mRS of 3 or greater had increased hazard for recurrent stroke in PRoFESS (hazard ratio [HR], 1.63; 95% CI, 1.38-1.92; P < .001) and in IRIS (HR, 1.91; 95% CI, 1.28-2.86; P = .002). There was also increased hazard for MACE in PRoFESS (HR, 1.90; 95% CI, 1.66-2.18; P < .001) and in IRIS (HR, 1.45; 95% CI, 1.03-2.03; P = .03). This cohort study found that higher baseline poststroke disability was associated with increased rates of recurrent stroke and MACE. Including more patients with greater baseline disability in stroke prevention trials may improve the statistical power and generalizability of these studies.

Development of a differential treatment selection model for depression on consolidated and transformed clinical trial datasets

Major depressive disorder (MDD) is the leading cause of disability worldwide, yet treatment selection still proceeds via “trial and error”. Given the varied presentation of MDD and heterogeneity of treatment response, the use of machine learning to understand complex, non-linear relationships in data may be key for treatment personalization. Well-organized, structured data from clinical trials with standardized outcome measures is useful for training machine learning models; however, combining data across trials poses numerous challenges. There is also persistent concern that machine learning models can propagate harmful biases. We have created a methodology for organizing and preprocessing depression clinical trial data such that transformed variables harmonized across disparate datasets can be used as input for feature selection. Using Bayesian optimization, we identified an optimal multi-layer dense neural network that used data from 21 clinical and sociodemographic features as input in order to perform differential treatment benefit prediction. With this combined dataset of 5032 individuals and 6 drugs, we created a differential treatment benefit prediction model. Our model generalized well to the held-out test set and produced similar accuracy metrics in the test and validation set with an AUC of 0.7 when predicting binary remission. To address the potential for bias propagation, we used a bias testing performance metric to evaluate the model for harmful biases related to ethnicity, age, or sex. We present a full pipeline from data preprocessing to model validation that was employed to create the first differential treatment benefit prediction model for MDD containing 6 treatment options.

A survey on UK researchers’ views regarding their experiences with the de-identification, anonymisation, release methods and re-identification risk estimation for clinical trial datasets

Background: There are increasing pressures for anonymised datasets from clinical trials to be shared across the scientific community. However, there is no standardised set of recommendations on how to anonymise and prepare clinical trial datasets for sharing, while an ever-increasing number of anonymised datasets are becoming available for secondary research. Our aim was to explore the current views and experiences of researchers in the United Kingdom about de-identification, anonymisation, release methods and re-identification risk estimation for clinical trial datasets. Methods: We used an online exploratory cross-sectional descriptive survey that consisted of both open-ended and closed questions. Results: We had 38 responses to invitation from June 2022 to October 2022. However, 35 participants (92%) used internal documentation and published guidance to de-identify/anonymise clinical trial datasets. De-identification, followed by anonymisation and then fulfilling data holders’ requirements before access was granted (controlled access), was the most common process for releasing the datasets as reported by 18 (47%) participants. However, 11 participants (29%) had previous knowledge of re-identification risk estimation, but they did not use any of the methodologies. Experiences in the process of de-identifying/anonymising the datasets and maintaining such datasets were mostly negative, and the main reported issues were lack of resources, guidance, and training. Conclusion: The majority of responders reported using documented processes for de-identification and anonymisation. However, our survey results clearly indicate that there are still gaps in the areas of guidance, resources and training to fulfil sharing requests of de-identified/anonymised datasets, and that re-identification risk estimation is an underdeveloped area.

Comparison of Synthetic Data Generation Techniques for Control Group Survival Data in Oncology Clinical Trials: Simulation Study.

Synthetic patient data (SPD) generation for survival analysis in oncology trials holds significant potential for accelerating clinical development. Various machine learning methods, including classification and regression trees (CART), random forest (RF), Bayesian network (BN), and conditional tabular generative adversarial network (CTGAN), have been used for this purpose, but their performance in reflecting actual patient survival data remains under investigation. The aim of this study was to determine the most suitable SPD generation method for oncology trials, specifically focusing on both progression-free survival (PFS) and overall survival (OS), which are the primary evaluation end points in oncology trials. To achieve this goal, we conducted a comparative simulation of 4 generation methods, including CART, RF, BN, and the CTGAN, and the performance of each method was evaluated. Using multiple clinical trial data sets, 1000 data sets were generated by using each method for each clinical trial data set and evaluated as follows: (1) median survival time (MST) of PFS and OS; (2) hazard ratio distance (HRD), which indicates the similarity between the actual survival function and a synthetic survival function; and (3) visual analysis of Kaplan-Meier (KM) plots. Each method's ability to mimic the statistical properties of real patient data was evaluated from these multiple angles. In most simulation cases, CART demonstrated the high percentages of MSTs for synthetic data falling within the 95% CI range of the MST of the actual data. These percentages ranged from 88.8% to 98.0% for PFS and from 60.8% to 96.1% for OS. In the evaluation of HRD, CART revealed that HRD values were concentrated at approximately 0.9. Conversely, for the other methods, no consistent trend was observed for either PFS or OS. CART demonstrated better similarity than RF, in that CART caused overfitting and RF (a kind of ensemble learning approach) prevented it. In SPD generation, the statistical properties close to the actual data should be the focus, not a well-generalized prediction model. Both the BN and CTGAN methods cannot accurately reflect the statistical properties of the actual data because small data sets are not suitable. As a method for generating SPD for survival data from small data sets, such as clinical trial data, CART demonstrated to be the most effective method compared to RF, BN, and CTGAN. Additionally, it is possible to improve CART-based generation methods by incorporating feature engineering and other methods in future work.

Uncertainty estimation using a 3D probabilistic U-Net for segmentation with small radiotherapy clinical trial datasets

Background and objectivesBio-medical image segmentation models typically attempt to predict one segmentation that resembles a ground-truth structure as closely as possible. However, as medical images are not perfect representations of anatomy, obtaining this ground truth is not possible. A surrogate commonly used is to have multiple expert observers define the same structure for a dataset. When multiple observers define the same structure on the same image there can be significant differences depending on the structure, image quality/modality and the region being defined. It is often desirable to estimate this type of aleatoric uncertainty in a segmentation model to help understand the region in which the true structure is likely to be positioned. Furthermore, obtaining these datasets is resource intensive so training such models using limited data may be required. With a small dataset size, differing patient anatomy is likely not well represented causing epistemic uncertainty which should also be estimated so it can be determined for which cases the model is effective or not. MethodsWe use a 3D probabilistic U-Net to train a model from which several segmentations can be sampled to estimate the range of uncertainty seen between multiple observers. To ensure that regions where observers disagree most are emphasised in model training, we expand the Generalised Evidence Lower Bound (ELBO) with a Constrained Optimisation (GECO) loss function with an additional contour loss term to give attention to this region. Ensemble and Monte-Carlo dropout (MCDO) uncertainty quantification methods are used during inference to estimate model confidence on an unseen case. We apply our methodology to two radiotherapy clinical trial datasets, a gastric cancer trial (TOPGEAR, TROG 08.08) and a post-prostatectomy prostate cancer trial (RAVES, TROG 08.03). Each dataset contains only 10 cases each for model development to segment the clinical target volume (CTV) which was defined by multiple observers on each case. An additional 50 cases are available as a hold-out dataset for each trial which had only one observer define the CTV structure on each case. Up to 50 samples were generated using the probabilistic model for each case in the hold-out dataset. To assess performance, each manually defined structure was matched to the closest matching sampled segmentation based on commonly used metrics. ResultsThe TOPGEAR CTV model achieved a Dice Similarity Coefficient (DSC) and Surface DSC (sDSC) of 0.7 and 0.43 respectively with the RAVES model achieving 0.75 and 0.71 respectively. Segmentation quality across cases in the hold-out datasets was variable however both the ensemble and MCDO uncertainty estimation approaches were able to accurately estimate model confidence with a p-value < 0.001 for both TOPGEAR and RAVES when comparing the DSC using the Pearson correlation coefficient. ConclusionsWe demonstrated that training auto-segmentation models which can estimate aleatoric and epistemic uncertainty using limited datasets is possible. Having the model estimate prediction confidence is important to understand for which unseen cases a model is likely to be useful.

Development and validation of MOSPROC, an easily applicable predictive score of mortality for patients with prostate cancer.

e13598 Background: Predicting life expectancy in prostate cancer patients is particularly difficult as different clinical factors significantly influence clinical outcome. Here we have developed MOSPROC, a simplified mortality predictive score to identify prostate cancer patients at high risk of mortality. Methods: The training set (TRS) (n=2,035) included metastatic Castration Resistant PC (mCRPC) patients from 4 randomized phase 3 trials (NCT00273338, NCT00988208, NCT00617669, and NCT00519285). The validation set included patients from NCT00554229 (mCRPC; n=256) (TS1) and NCT00626548 (M0 CRPC; n=661) (TS2) trials. Clinical trial datasets were obtained from www.projectdatasphere.org . All datasets contained more than 100 clinical variables per patient at baseline. Synthetic representations of every patient were generated and input into a deep learning framework to identify subgroup of patients based on their similarities. The resultant subgroups were correlated with overall survival (OS). Differential variables between subgroups were identified and included in a multivariable logistic regression model. Independent predictive factors found to be statistically significant were used to generate a predictive score of survival that was validated in the validation sets. Results: The deep learning framework identified two different subpopulations in the TRS: LL (n=824) and HH (n=1,204) with significantly different survival outcome. Patients in HH had a higher risk of death compared to LL (median OS 17.6 months vs. 27.3 months, respectively; HR 2.09 95%CI 1.84-2.39; p&lt;.0001). Feature contribution analysis identified 36 differential features between both subpopulations including anthropometrics, performance status, vital signs, biochemistry, previous and current cancer therapies, and concomitant medications. From these, PSA, ALP, and AST were the best independent predictive factors that were then selected to create a simplified predictive score for survival. A score ≥34 points was considered a high score (HS) while if &lt; 34, a low score (LS). When applied to TRS, the predictive score yielded an area under the curve of 0.91. Patients with HS presented a significantly shorter OS (median OS 18.4 vs. 23.9 months; HR 1.49 95%CI 1.31-1.71; p&lt;.005). The predictive score was then validated in two additional trials with similar results. In both cases, patients with a HS presented a significantly shorter OS (TS1: HR 2.56, 95%CI 1.79-3.66, p&lt;.01; TS2: HR: 3.46, 95%CI 1.89-6.33, &lt;.001). Conclusions: By combining deep learning and classical biostatistical tools, we have developed MOSPROC, an easily applicable predictive score that, based on baseline values of PSA, ALP and AST, can identify prostate cancer patients with a higher risk of death. Further work is required to validate this approach in prospective cohorts of PC patients in additional clinical trials and/or real-world datasets.

Real-world and clinical trial validation of a deep learning radiomic biomarker for PD-(L)1 immune checkpoint inhibitor response in stage IV NSCLC.

102 Background: Immune checkpoint inhibitor (ICI) therapy is standard-of-care for treatment of mutation-negative advanced non-small cell lung cancer (NSCLC). However, given known inaccuracy of PD-(L)1 markers, there is an unmet need to better identify patients most likely to derive clinical benefit from ICI. We developed and externally validated a generalizable CT imaging-based biomarker to predict response to ICI. Methods: We developed and validated a deep learning radiomic biomarker using an internally curated real-world dataset (RWD) of 2,010 stage IV NSCLC patients treated with PD-(L)1 ICIs in academic and community settings from US and Europe. Patients with missing baseline imaging, missing follow-up data, or EGFR/ALK oncogenic driver mutations were excluded, resulting in a total of 1,188 subjects. This RWD consisted of a discovery cohort (Dataset A, N=844) and a temporally distinct holdout cohort (Dataset B, N=344), which were used to generate performance metrics of the biomarker. To test generalizability, we validated our biomarker in a prospective clinical trial dataset evaluating Sasanlimab in PD-(L)1 therapy-naïve, advanced NSCLC patients (NCT02573259, Dataset C, N=54). We utilized a two-stage learning approach to model 6-month PFS. First, we used an independent multi-task deep-learning feature extractor trained on 19,184 whole chest CT scans. Second, we input the extracted features into a Cox proportional hazard (CoxPH) model along with age, sex, and baseline lesion measurements (sum of longest diameter and distant metastases counts), and the model generated a time-dependent PFS function and a response score. We performed 6-fold cross-validation on Dataset A to train and evaluate the models, which were subsequently ensembled and applied to independent Datasets B and C. To assess independence from PD-L1 status and key demographic covariates, we herein report multivariate adjusted hazard ratios (HR) for the group identified as low-risk based on the biomarker. Results: In Dataset A, the biomarker showed a cross-validation PFS adjusted HR of 0.49 (95% CI 0.38-0.63) in the all-comers cohort and 0.28 (0.17-0.46) in the first-line ICI monotherapy cohort (1LMono). In Dataset B, the PFS adjusted HRs were 0.54 (0.35-0.83) in all-comers and 0.18 (0.05-0.61) in 1LMono. In Dataset C, the adjusted HRs were 0.30 (0.14-0.68) for PFS, 0.29 (0.10-0.83) for OS, 0.31 (0.14-0.72) for TTP. Conclusions: In our validations in RWD and clinical trial cohorts, a deep-learning radiomic biomarker based on routine pre-treatment CT scans predicted response to ICI and stratified patients independently from PD-L1 status. This tool may inform clinical decision-making, such as to help guide whether concomitant chemotherapy may not be needed. In future work, we plan to further validate our approach in larger prospective datasets and expand its use to new indications.

TERT-associated DNA polymerases genes link CAF and CD8+ T Cells to improve immunotherapy response rate across multiple cancers.

e14669 Background: The cancer cells exhibit heightened basal levels of responses to DNA breaks, accumulating DNA polymerases family (DNA-pol) as repair protein, thereby increasing DNA mutation and neoantigens. Based on this, we investigated the potential of DNA-pol as an immunotherapy biomarker and its underlying mechanism. Methods: We conducted a pan-cancer analysis of DNA-pol gene dysregulation in 6366 samples. Normal fibroblast, cancer-associated fibroblast, and CD8+ T cells were cultured from cancer patients (3 esophagus, 5 lung, 5 renal, and 3 liver cancer), then treated them with PD-L1 inhibitor or overexpressing TERT for RNA-seq and proteomics. 996 multiple cancer cases with ICIs were collected to test biomarker efficiency. A phase II clinical trial is evaluating and exploring our biomarker in mechanism prospectively. Results: We defined a distinct signature score of DNA-pol genes upregulated in cancer (DNA-pol-up) and linked to worse prognosis (HR=1.2). DNA-pol-up program dysregulation led to immune processes alteration and was related to upstream activation of TERT signaling (z-score=3.6) in ingenuity pathway analysis and linked to immunologically active tumors by recruiting CD8+ T cell and downregulating CAF. DNA-pol-up score was elevated in PD-L1-treated CAF and TERT-overexpressed CD8+ T cell, in MSI-H cancer, also related to more neoepitope peptides, TMB (Rho=0.54), cancer stemness (Rho=0.78), and tumor purity (Rho=0.29). Cancers that activate this program carried distinct genomic profiles with PIK3CA, KRAS, and TP53 mutations. This signature was an independent predictor for ICIs response (OR=2.52), even outperforming cytolytic activity, MSI, TMB (OR=0.9). In 12 clinical trial datasets, this signature could stratify the responding patients (maxAUC=0.762). In our phase II clinical trial, 34 liver cancer were enrolled with 4-year follow-up, it also has a satisfactory performance in predicting the ORR (AUC=0.699) and classifies mortality rate (HR=2.5), meanwhile verified its relationship with detailed potential molecular mechanism. Conclusions: Our findings identify a distinct transcriptional pattern of DNA-pol genes across cancers, which highlighted the role of DNA-pol family genes in predicting the immunotherapy response for the first time, and TERT could be a novel vaccine candidate for improving immunotherapy response.

Deep Learning-Based Automatic Classification of Ischemic Stroke Subtype Using Diffusion-Weighted Images.

Accurate classification of ischemic stroke subtype is important for effective secondary prevention of stroke. We used diffusion-weighted image (DWI) and atrial fibrillation (AF) data to train a deep learning algorithm to classify stroke subtype. Model development was done in 2,988 patients with ischemic stroke from three centers by using U-net for infarct segmentation and EfficientNetV2 for subtype classification. Experienced neurologists (n=5) determined subtypes for external test datasets, while establishing a consensus for clinical trial datasets. Automatically segmented infarcts were fed into the model (DWI-only algorithm). Subsequently, another model was trained, with AF included as a categorical variable (DWI+AF algorithm). These models were tested: (1) internally against the opinion of the labeling experts, (2) against fresh external DWI data, and (3) against clinical trial dataset. In the training-and-validation datasets, the mean (±standard deviation) age was 68.0±12.5 (61.1% male). In internal testing, compared with the experts, the DWI-only and the DWI+AF algorithms respectively achieved moderate (65.3%) and near-strong (79.1%) agreement. In external testing, both algorithms again showed good agreements (59.3%-60.7% and 73.7%-74.0%, respectively). In the clinical trial dataset, compared with the expert consensus, percentage agreements and Cohen's kappa were respectively 58.1% and 0.34 for the DWI-only vs. 72.9% and 0.57 for the DWI+AF algorithms. The corresponding values between experts were comparable (76.0% and 0.61) to the DWI+AF algorithm. Our model trained on a large dataset of DWI (both with or without AF information) was able to classify ischemic stroke subtypes comparable to a consensus of stroke experts.

ClinicalGAN: powering patient monitoring in clinical trials with patient digital twins

Conducting clinical trials is becoming increasingly challenging lately due to spiraling costs, increased time to market, and high failure rates. Patient recruitment and retention is one of the key challenges that impact 90% of the trials directly. While a lot of attention has been given to optimizing patient recruitment, limited progress has been made towards developing comprehensive clinical trial monitoring systems to determine patients at risk and potentially improve patient retention through the right intervention at the right time. Earlier research in patient retention primarily focused on using deterministic frameworks to model the inherently stochastic patient journey process. Existing generative approaches to model temporal data such as TimeGAN or CRBM , face challenges and fail to address key requirements such as personalized generation, variable patient journey, and multi-variate time-series needed to model patient digital twin. In response to these challenges, current research proposes ClinicalGAN to enable patient level generation, effectively creating a patient’s digital twin. ClinicalGAN provides capabilities for: (a) patient-level personalized generation by utilizing patient meta-data for conditional generation; (b) dynamic termination prediction to enable pro-active patient monitoring for improved patient retention; (c) multi-variate time-series training to incorporate relationship and dependencies among different tests measures captured during patient journey. The proposed solution is validated on two Alzheimer’s clinical trial datasets and the results are benchmarked across multiple dimensions of generation quality. Empirical results demonstrate that the proposed ClinicalGAN outperforms the SOTA approach by 3–4×\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ imes $$\\end{document} on average across all the generation quality metrics. Furthermore, the proposed architecture is shown to outperform predictive methods at the task of drop-off prediction significantly (5–10% MAPE scores).

Competing risks and multivariate outcomes in epidemiological and clinical trial research.

Data analysis methods for the study of treatments or exposures in relation to a clinical outcome in the presence of competing risks have a long history, often with inference targets that are hypothetical, thereby requiring strong assumptions for identifiability with available data. Here data analysis methods are considered that are based on single and higher dimensional marginal hazard rates, quantities that are identifiable under standard independent censoring assumptions. These lead naturally to joint survival function estimators for outcomes of interest, including competing risk outcomes, and provide the basis for addressing a variety of data analysis questions. These methods will be illustrated using simulations and Women's Health Initiative cohort and clinical trial data sets, and additional research needs will be described.

A clinical trial termination prediction model based on denoising autoencoder and deep survival regression

AbstractEffective clinical trials are necessary for understanding medical advances but early termination of trials can result in unnecessary waste of resources. Survival models can be used to predict survival probabilities in such trials. However, survival data from clinical trials are sparse, and DeepSurv cannot accurately capture their effective features, making the models weak in generalization and decreasing their prediction accuracy. In this paper, we propose a survival prediction model for clinical trial completion based on the combination of denoising autoencoder (DAE) and DeepSurv models. The DAE is used to obtain a robust representation of features by breaking the loop of raw features after autoencoder training, and then the robust features are provided to DeepSurv as input for training. The clinical trial dataset for training the model was obtained from the ClinicalTrials.gov dataset. A study of clinical trial completion in pregnant women was conducted in response to the fact that many current clinical trials exclude pregnant women. The experimental results showed that the denoising autoencoder and deep survival regression (DAE‐DSR) model was able to extract meaningful and robust features for survival analysis; the C‐index of the training and test datasets were 0.74 and 0.75 respectively. Compared with the Cox proportional hazards model and DeepSurv model, the survival analysis curves obtained by using DAE‐DSR model had more prominent features, and the model was more robust and performed better in actual prediction.

Segmentation of patients with small cell lung cancer into responders and non-responders using the optimal cross-validation technique

BackgroundThe timing of treating cancer patients is an essential factor in the efficacy of treatment. So, patients who will not respond to current therapy should receive a different treatment as early as possible. Machine learning models can be built to classify responders and nonresponders. Such classification models predict the probability of a patient being a responder. Most methods use a probability threshold of 0.5 to convert the probabilities into binary group membership. However, the cutoff of 0.5 is not always the optimal choice.MethodsIn this study, we propose a novel data-driven approach to select a better cutoff value based on the optimal cross-validation technique. To illustrate our novel method, we applied it to three clinical trial datasets of small-cell lung cancer patients. We used two different datasets to build a scoring system to segment patients. Then the models were applied to segment patients into the test data.ResultsWe found that, in test data, the predicted responders and non-responders had significantly different long-term survival outcomes. Our proposed novel method segments patients better than the standard approach using a cutoff of 0.5. Comparing clinical outcomes of responders versus non-responders, our novel method had a p-value of 0.009 with a hazard ratio of 0.668 for grouping patients using the Cox proportion hazard model and a p-value of 0.011 using the accelerated failure time model which approved a significant difference between responders and non-responders. In contrast, the standard approach had a p-value of 0.194 with a hazard ratio of 0.823 using the Cox proportion hazard model and a p-value of 0.240 using the accelerated failure time model indicating the responders and non-responders do not differ significantly in survival.ConclusionIn summary, our novel prediction method can successfully segment new patients into responders and non-responders. Clinicians can use our prediction to decide if a patient should receive a different treatment or stay with the current treatment.

Deep Reinforcement Learning for Early Diagnosis of Lung Cancer

Lung cancer remains the leading cause of cancer-related death worldwide, and early diagnosis of lung cancer is critical for improving the survival rate of patients. Performing annual low-dose computed tomography (LDCT) screening among high-risk populations is the primary approach for early diagnosis. However, after each screening, whether to continue monitoring (with follow-up screenings) or to order a biopsy for diagnosis remains a challenging decision to make. Continuing with follow-up screenings may lead to delayed diagnosis but ordering a biopsy without sufficient evidence incurs unnecessary risk and cost. In this paper, we tackle the problem by an optimal stopping approach. Our proposed algorithm, called EarlyStop-RL, utilizes the structure of the Snell envelope for optimal stopping, and model-free deep reinforcement learning for making diagnosis decisions. Through evaluating our algorithm on a commonly used clinical trial dataset (the National Lung Screening Trial), we demonstrate that EarlyStop-RL has the potential to greatly enhance risk assessment and early diagnosis of lung cancer, surpassing the performance of two widely adopted clinical models, namely the Lung-RADS and the Brock model.

Abstract 2282: Predictive association of tumor cells hijacked NFATC2 activities with resistance to immune checkpoint inhibition in melanoma patients

Abstract Background: Immune checkpoint blockade (ICB) has emerged as an effective cancer therapy. However, only a small fraction of patients receives benefits whereas a significant subset of patients experienced severe or even lethal autoimmune diseases. Stratifying patients that would benefit from immunotherapy remain a highly unmet clinical need. Dysregulations of transcriptional programs have been implicated in cancer initiation and progression. To date, there is a paucity of studies on the predictive values of transcriptional factors on ICB response based on the most recently emerged ICB clinical trial datasets. We reasoned that cancer cells may hijack key transcriptional programs in hematological cells to evade the attack from immune cells and acquired tolerance to immunotherapy. Method: Herein, we used an R package singscore/1.14.0 to compute the TFs activity scores based on their validated gene targets (TGs). Leveraging melanoma clinical trial datasets, we analyzed the predictive association of the TF scores with resistance to anti-PD1 ICB. The R package, iGenSig was used to evaluate the potential confounding variables. Results: Our landscape correlation analysis revealed NFATC2 (Nuclear Factor Of Activated T Cells 2)-TGs as the most significant predictor of anti-PD-1 ICB resistance in melanoma. NFATCT-TGS as a predictor of anti-PD1 resistance is confounded by liver metastasis and achieved AUC range between 0.71 to 0.74 in three datasets (PRJEB23709, GSE160638, and Phs000452). In addition, prior anti-CTLA4 exposure (AUC=0.87 and 0.89) is associated with higher predictive efficacy of NFATC2-TGS than anti-CTLA4 naïve tumor (AUC=0.61 and 0.69) in phs000452 and GSE91061 datasets. Survival analysis revealed that NFATC2 is a better predictor of progressive-free survival (PFS; HR=5.95), and overall survival (OS; HR= 3.28-3.38) in anti-CTLA4 exposed tumor than anti-CTLA4 naïve tumor (PFS: HR=1.92, and OS; HR=1.64-2.17). Finally, NFATC2-TGS show strong inhibitory correlation with B cells and T cells infiltration, but strongly correlated with tumor purity. Conclusion: We demonstrate that NFATC2 TGS predicts ICB benefits in melanoma patients, and thus may guide the recruitment of melanoma patients for anti-PD1 ICB. Our data provide support to the hypothesis that melanoma acquired NFATC2 transcriptional program to avoid immune cells evasion and acquired tolerance to ICB immunotherapy Citation Format: Bashir Lawal, Yue Wang, Parisa Lotfinejad, Xiaosong Wang. Predictive association of tumor cells hijacked NFATC2 activities with resistance to immune checkpoint inhibition in melanoma patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 2282.