Predicting Individual Responses in Phase I Oncology Trials Using Routinely Collected Clinical Biomarkers

doi:10.21203/rs.3.rs-4910461/v1

Information which may support an individual’s participation in a cancer phase I trial, such as their response to prior therapies, other medical conditions they may have, features in their tumor genomic profile, etc., should be considered to avoid negative consequences of participating in the trial. However, knowing which pieces of information are relevant is crucial. We built predictive models of responses in phase I trials using routinely collected demographic and clinical chemistry data. We obtained data on 1386 participants in 252 phase I trials pursued within the US Oncology clinical trial network in the years 2007–2018. We used mixed models, as well as machine learning (ML) techniques exploiting randomly generated training and test data sets, to build predictive models of four different outcomes while controlling for which trial a participant was enrolled in. The outcomes of interest were time on the experimental therapy, time on the study drug relative to the drug during which their cancer progressed, 90-day mortality, and Grade 5 toxicity. We also controlled for other potential sources of variation in outcomes such as weight, height, and sex. We found that an unfavorable participant profile includes elevated white blood cells, low albumin levels, and low hemoglobin levels, as well as low BMI for mortality risk, among other factors, many of which are consistent with previously published findings. In addition, our ML-based predictions achieved, on average, > 80% area under the receiver/operator curve (AUC) statistics reflecting good accuracy for predicting dichotomous outcomes. Our findings could be of general use when recruiting for Phase I oncology clinical trials.

Treatment Response

Clinical Trial

Biomarkers

Drug Development

Phase I Studies

Machine Learning

Clinical Predictions

When an individual with cancer does not respond to their current medication, questions arise as to what additional medications or interventions should be considered to treat their cancer and improve their health. If alternatives and additional treatments are not obvious, or if many different interventions have not worked in the past, then enrollment in a clinical trial exploring the benefits of a novel therapy make sense [1–7]. The choice of which trial to enroll in can be greatly enhanced by leveraging, e.g., genomic sequencing and related tumor and treatment history profiling strategies. Such profiling is consistent with the growing adoption and use among oncologists of technologies that enable personalization of interventions, such as tumor sequencing [8–12]. Beyond, e.g., tumor genomic profiling and drug target information, any information that would help differentiate interventions likely to benefit – and certainly not harm – a participant would be ideal to know [8–12]. This raises the question of whether other factors, such as a participant’s general characteristics, their health trajectory, and their previous experience with interventions, can be used to determine the likely benefits of any novel or experimental therapy. Lack of insight into what might benefit participants has led to recent concerns about the personal benefit of participants in clinical trials.[13–15]

Many candidate biomarkers of response in early Phase trials (e.g., IA and IB; which we refer ‘Phase I’ studies; see Supplementary Table 1), including routinely collected clinical chemistries, tumor parameters, and demographic variables such as albumin levels, lactate dehydrogenase (LDH), alkaline phosphatase (ALP), the number of metastatic sites, white blood cell count, and anemia.[16–18]. Unfortunately, many of the previously published studies exploring predictive models for Phase I responses are limited. Some studies have small sample sizes, and some have a limited number of potential predictive markers that were considered. However, all previous studies were limited in that they focused on data from participants in a wide variety of trials where the agents being tested likely differed in efficacy. One might expect to see poorer outcomes in studies where the drugs were ultimately found not to be efficacious (or studies of interventions less likely to create adverse effects). Thus, a search for prognostic biomarkers of overall response is important over and above the search for markers that are predictive of response to a particular intervention [17, 19–27].

To further identify and evaluate potential prognostic markers of benefit in Phase I clinical trials, we gathered data on 2362 participants, 1386 with enough data for analysis, in 252 Phase I trials within the US Oncology clinical trials network between 2007–2018. We considered four Phase I trial outcomes for our analyses: time on study (for which we are making the assumption that participants staying on the intervention in a Phase I study are benefitting from that intervention to some degree); the ratio of time a participant spent on Phase I agent relative to the time the spent on the therapy whose ineffectiveness motivated their participation in a Phase I study [28–30]; 90-day mortality (treated as a yes/no outcome); and overall Grade 5 toxicity (i.e., death; a yes/no outcome) during the trial. Note that we did not have length of overall survival (OS) data post-trial completion given the complexities of obtaining that data across the individual clinics contributing to the studies we drew on. We considered many of the routine clinical chemistries previous researchers explored in their analyses (Supplementary Table 1) in addition to a few others (see Supplementary Tables 1 and 2). Given the number of participants in each trial we had for our analyses, we controlled for variation across the individual trials using random effects models [31, 32], to accommodate potential differential efficacy but compared our results to emerging artificial intelligence (AI) and machine learning (ML) techniques. We provide example predictions but recognize that generalizability to different settings may not be warranted.

A more complete description of the data sources and analytical methods used, including specification of technical parameters used in the AI/ML analyses are provided in the Supplementary Methods section.

Participant Data Sources

We gathered de-identified information from the US Oncology Network (USON) Electronic Health Record (EHR), known as iKnowMed (iKM), and the Limited Access Death Master File (LADMF), associated with USON trial participants who: 1. Participated in USON phase IA and phase IB clinical trials between 4/2007-6/2018, with follow-up through 12/2018 or until date of the participant’s last record. These trials include USON’s phase IA (single agent set) and phase IB (combination phase IB data set) trials; 2. Received care at a USON site utilizing the full EHR capacities of iKnowMed (iKM) at the time of treatment; and 3. Were ≥ 18 years of age at trial participation. Any participant who had previously opted out of their results being kept in the USON database were not included. The LADMF, which was our principal source of vital status (death) of the participating participants, is overseen by the Social Security Administration (SSA). To accommodate missing data, we imputed values for missing variables using the Multiple Imputation with Chained Equation (MICE) method [33]. To determine how sensitive our results were to imputation, we pursued sensitivity analyses in which we compared analysis results generated on both the complete and imputed data sets.

Responses and Candidate Prognostic Biomarkers

We considered 4 different outcomes, or dependent variables, in our analyses: 1. The duration of time a participant was on the study drug, which is a surrogate measure for whether the drug was working and for how long it appeared to be working; 2. The amount of time a participant was on the Phase I study drug divided into the time on the previous drug during which their cancer progressed (i.e., a smaller value is better for the Phase I drug experience). This measure is reflected as a ratio and gives an indication of how well the participant is responding to the study drug relative to a drug that was not benefitting them [28–30]; 3. 90-day mortality (i.e., whether a participant dies within 90 days of the initiation of the trial, treated as a yes/no dichotomous variable); and 4. Grade 5 toxicity or death at any time during the trial (also treated as yes/no dichotomous variable). We did not have overall survival across the large and diverse US Oncology system, especially after trials were completed. We considered many independent variables in our analyses (see also Supplementary Methods and Tables 2 and 3). All continuous variables (e.g., BMI) were standardized to have zero mean and unit standard deviation. Categorical variables were recoded as dummy variables (e.g., sex: 0 = female, 1 = male).

Table 1: Trial indices that were significant predictors (p<0.05) of response across the 252 Phase I trials based on stepwise logistic regression for the dichotomous outcomes.

90-day Mortality				Grade 5 Toxicity
Trial Index	Coef	Std	p-value	Trial Index	Coef	Std	p-value
51	2.733	0.840	0.001	23	2.298	0.687	<0.001
0	1.016	0.270	<0.001	0	1.467	0.393	<0.001
7	1.063	0.436	0.015	41	2.298	0.833	0.006
50	1.529	0.768	0.047	38	2.298	0.833	0.006
3	-1.441	0.725	0.047	31	2.144	0.819	0.009
				1	1.184	0.431	0.006

Key: Each entry provides the trial index (1-190), number of subjects in the trial, regression coefficient+/-standard error, and p-value; Time on therapy = time on the experimental therapy; Relative time = time on the study agent relative to time on the agent during which the participant’s cancer progressed.

Table 2

a. Final Mixed Effect Models for the Continuous Outcomes.
Time on Treatment				Relative Time on Treatment
Biomarker	Coeff	SE	p-val	Biomarker	Coeff	SE	p-val
Hemoglobin	11.876	5.287	0.025	DOPT	2.123	0.251	7.97E-17
WBC	-14.517	5.409	0.007	WBC	2.240	0.881	0.012
Lymphocytes	13.542	6.828	0.057	Albumin	-0.613	0.255	0.016
LDH	-15.834	4.956	0.001	Uric acid	0.751	0.248	0.003
Albumin	10.705	5.371	0.046
Breast Cancer	33.460	16.808	0.047
ECOG score	-51.145	18.645	0.006

Table 2

b. Final Mixed Effect Models for the Dichotomous Outcomes.
90-day mortality				Grade 5 toxicity
Biomarker	Coeff	SE	p-val	Biomarker	Coeff	SE	p-val
WBC	0.539	0.082	4.35E-11	Albumin	-0.476	0.194	0.014
Albumin	-0.354	0.113	0.001	WBC	0.483	0.140	0.0005
LDH	0.260	0.072	0.0002	Hemoglobin	-0.324	0.758	0.670
Hemoglobin	-0.281	0.095	0.003	ECOG score	1.218	0.521	0.019
BMI	-0.228	0.091	0.0125
DOPT	-0.355	0.126	0.005
Lymphocytes	-0.331	0.125	0.008

Table 3

Significance of the inter-trial heterogeneity based on treating the y-intercept for each prognostic biomarker and outcome as a random effect.
Prognostic Factor	Time on Therapy	Relative Time	90-day Mortality	Grade 5 Toxicity
Albumin	5.897936e-21	3.340707e-30	0.078	1.169658e-06
WBC	1.341636e-23	2.961202e-79	NA	1.218731e-06
LDH	2.45875e-21	NA	NA	NA
Hemoglobin	NA	NA	NA	6.569533e-07
Lymphocytes	3.733162e-25	NA	0.066	NA

Key: Each entry provides the p-value testing the significance of variation in the y-intercept across the trials with respect to each prognostic factor and each outcome; Time on therapy = time on the experimental therapy; Relative time = time on the study agent relative to time on the agent during which the participant’s cancer progressed; WBC = White Blood Cells; LDH = Lactate dehydrogenase. NA = Not applicable since the prognostic factor did not exhibit significant or near significant random effects.

Analytical Methods

We used standard descriptive statistics (mean, standard deviation, frequencies, etc.) to characterize the responses and candidate prognostic biomarkers. We also used standard Pearson parametric and Spearman non-parametric correlation analyses to assess the basic relationships between the biomarkers. We focused our analyses on mixed effects (or random effects) linear regression and logistic regression models to control for variation in biomarker and response relationships across the trials [31, 32]. To compare results with and without accounting for inter-trial variation, we also used standard linear and logistic regression, including stepwise linear and logistic regression. We also used survival analyses to analyze time on trial response [34]. Trial index (a number indicating the trial associated with each participant) was included as a random effect factor in these mixed model analyses. We also considered the trial indices as fixed effects in initial analyses to assess each trial’s association with the responses relative to the other trials. We note that many of the trials we drew data from were only represented by 1 or few participants. To carry out the mixed model analyses we used the modules ‘lme4’, ‘GLMMAdaptive,’ and ‘Coxph,’ in the R computing environment [35].

We used many well-known ML methods for the analysis of the continuous outcome variables (i.e., Time on Study and Time on Prior Therapy Ratio) as implemented in the Sci-kit learn package [36]. These methods are discussed in many excellent references [37]. As a way of aggregating the results of the different ML methods to build an ensemble model, we used the ‘Super Learner’ (SL) technique [38, 39]. For the ML analyses we considered trial indices as fixed effects, but only for those trials that were significant in the initial stepwise regression analysis and that had enough participants. In addition, the Supplementary Methods section also describes how we handled imbalances in the data for the ML models using the ‘Synthetic Minority Over-sampling Technique (SMOTE)’ [40–42], how we evaluated the performance of each model using standard metrics [37, 43] and Cohen’s Kappa [40] for area-under-the-curve (AUC) from receiver-operator characteristics (ROC) curve analysis, and how we explored the sensitivity of our modeling. Finally, we used some example participant profiles to show how the development of our mixed effects models’ results can be used to make individual predictions of response in Phase I studies for the categorical outcome. Logistic regression models corrected for the imbalance in the categorial outcomes using the SMOTE procedure (see above and the Supplementary Methods) were used for these example predictions.

Final Data Set

The variables used in the analyses and the number of participants with values for those variables are listed in Supplementary Table 2 and complete descriptive statistics for each quantitative variable are provided in Supplementary Table 3. In addition, the frequencies of different tumor types among those identified as participating in the Phase I trials in our data set are provided in Supplementary Table 4. Overall, males made up 51.1% of our sample, 58.6% participants identified as white, and the ECOG scores ranged from 0 = 11.4% of participants in our study, 1 = 69.4%, 2 = 7.3% and 3 < 1%. The average (standard deviation), median, and mode of the number participants in these trials were 5.42 (9.67), 2.0, and 1.0 respectively. 30 trials had more than 10 participants and 102 had only 1 participant. Figure 1 provides correlations between the candidate continuous prognostic biomarkers with Standard Pearson (below the diagonal) and Spearman non-parametric (above the diagonal).

Initial Trial Index and Tumor Type Analysis

We initially explored the relationships between the trial indices and the outcomes. Table 1 provides the results of a stepwise logistic regression analysis for the two dichotomous outcomes and Supplementary Table 5 has the stepwise regression results of the two continuous outcomes. Table 1 suggests that 5 or 6 of the trials have a stronger association with the outcomes than others, suggesting that some trials may have tested, e.g., inferior therapies relative to others, or enrolled participants that may have been more compromised to begin with and more likely to manifest, e.g., negative outcomes of interest. These analyses also suggest there is potential for heterogeneity in the relationships between the prognostic biomarkers and the outcomes that could possibly be accommodated in random effects models.

The results of the multiple regression and stepwise regression results relating the candidate prognostic factors and the different cancer types to the outcomes are provided in Supplementary Table 6a for the continuous outcomes using linear regression (as well treating time on study as a survival endpoint in a survival analysis), and Supplementary Table 6b for the dichotomous outcomes using logistic regression. These analyses did not account for random effects across the trials. Hemoglobin and albumin both had positive relationships with time on study on, as did whether the trial was focused on renal cell carcinoma, although if the trial was focused on rectal cancer it had a negative relationship based on stepwise multiple regression [37]. For time on therapy ratio, no cancer type, but DOPT score, WBC, and Uric Acid had positive effects. For the survival analysis of time on study, only albumin had a significant, yet negative relationship. Due most likely to multicollinearity, a few other variables had significant relationships with the continuous outcomes from a standard multiple linear regression analysis that included all candidate prognostic factors. For the dichotomous outcomes, many more candidate prognostic variables exhibited associations, with hemoglobin levels and albumin exhibiting negative relationships and WBC and LDH exhibiting positive relationships with both outcomes based on the stepwise multiple logistic analysis. Renal cell carcinoma and non-Hodgkin’s lymphoma trials were negatively associated with both outcomes, while liver cancer trials were positively associated based on stepwise multiple logistic regression.

Machine Learning Model Analysis

Analyses evaluating the performance of the various ML models based on many widely used metrics (see Methods section) are provided in Supplementary Tables 7–9 for each outcome. We emphasize that the performance metrics for the models are based on analyses of the test data sets, not the training data sets. More detailed information including a description of the various performance metrics for each model is provided in the Supplementary Methods section. In particular, to overcome imbalances termed in the number of individuals exhibiting the mortality phenotypes during the trials versus those that did not, we used the ‘Synthetic Minority Over-sampling Technique (SMOTE)’ [40, 41] (see the Supplementary Methods). In addition, the ML methods did not account for random effect variation across the trials, but rather focused on the overall accuracy of the predictions about outcomes. Therefore, individual variable importance was not an emphasis in these analyses. However, Supplementary Fig. 1 provides a graphical representation of the importance or influence of each independent variable on 90-day mortality and Grade 5 toxicity based on the Random Forest (RF) analyses and suggests that albumin levels and WBC and whether the therapy tested was a targeted therapy are more important predictors of response. Figure 2 provides example ROC curve plots derived from the different ML methods for these two categorical outcomes. Note that some techniques produced the same ROC curves, so are overlapping. Ultimately, our use of an aggregated model, the Super Learner, provided some of the best fits (Fig. 2), which makes sense given that it pulls together information from the other models.

Mixed Model Analysis

As noted, we used trial index as a grouping variable in mixed or random effects models to account for variation across the trials (see Methods). Final models only considered significant random intercept effects based on univariate model analysis for the various outcomes. The most significant predictors are described in Table 2a for the continuous outcomes and 2b for the dichotomous outcomes. No random effects survival analysis was pursued for the time on treatment outcome. Table 3 provides the significance levels for those predictors that exhibited significant (p < 0.05) random effects across the outcomes in univariate models as well as exhibiting significant fixed effects from the stepwise multiple linear and logistic regression analyses. Figure 3 provides an example plot of the random effects for lymphocyte count and 90-day mortality assuming a y-intercept random effect and a slope effect in the model as clearly shows the relationship between lymphocyte count and 90-day mortality varies across the trials.

Time on Study. The strongest predictors for time on study include HGB, WBC, lymphocyte count, LDH, albumin level, ECOG2 assessment and Breast cancer (Left panels of Table 2a). WBC, LDH, and ECOG2 assessment, exhibited negative relationships and the others had positive relationships. All but hemoglobin exhibited significant random intercept effects (Table 3, first panel).

Time on Therapy Ratio. DOPT, WBC, albumin, and uric acid, were associated with Time on therapy ratio (Right panel of Table 2a). Only albumin and whether the therapy was targeted had negative effects and only albumin and WBC exhibited random effects (Table 3, 2nd panel).

90-Day Mortality. WBC, albumin, LDH, hemoglobin, BMI, DOPT and lymphocyte count were associated with 90-day mortality (left columns of Table 2b). Albumin, hemoglobin, BMI, DOPT, and lymphocyte count all had negative associations. BMI and DOPT did not exhibit random effects, but all the other predictor variables did (3rd column of Table 3).

Grade 5 Toxicity. Albumin, WBC, hemoglobin, and ECOG2 assessment were associated with grade 5 toxicity with albumin and hemoglobin exhibiting negative associations (right panel of Table 2b). Albumin, WBC and hemoglobin also exhibited random effects (last column of Table 3).

Sensitivity Analysis

We considered the potential influence of the use of imputed data on our results [33]. Supplementary Fig. 2 provides the results and suggests there was no overt difference as a comparison of models using participants with and without imputed values did not reveal any notable and consistent differences based on the various metrics for accuracy. Supplementary Table 10 has the results for the 90-day mortality analyses. Results for the other outcomes are not shown but are available upon request.

Example Predictions

We pursued example predictions based on the fixed effects from the final random effects models (Tables 2a and 2b; see Methods). Supplementary Table 11 provides the values used for the predictions as well as the coefficients used in the linear and logistic regression models. The code to produce the predictions is available from the authors but the use of the coefficients provided in Table 10 should be straightforward.

Summary of Findings

We find strong evidence for associations between routinely collected clinical chemistries, demographic information and clinically meaningful outcomes in Phase I trials, not unlike previous research (Supplementary Table 1). Although different sets of biomarkers exhibited associations with the different outcomes there were many prognostic biomarkers that were common to models for different outcomes and that had been reported in previous publications. Based on this previous and our current analyses, an unfavorable participant profile would include elevated white blood cells, low albumin levels, and low hemoglobin levels as well as low BMI for mortality risk. We were also able to create ML models that exhibited good predictive capacity that considered the combined effect of all our candidate prognostic biomarkers. We were careful not to let overfitting and missing data complicate our ML modeling and results by leveraging procedures for imputing missing data, using test and training sets to develop our models, and pursuing sensitivity and heterogeneity analyses to assess potential biases. Finally, we used different modeling techniques to see if they provided similar prediction accuracies. We ultimately found that this was not the case, and that the best performing models, on average, were the Super Learner models, which are aggregates of individual models. In addition, unlike other studies we were careful to consider variation across the trials with respect to the relationships they may exhibit between potential prognostic biomarkers and the outcomes using mixed (random) effects models. We found highly significant random effects associated with the trials, either focusing on y-intercept terms or slopes quantifying the relationships between the prognostic biomarkers and outcomes, suggesting heterogeneity across the studies. We believe this heterogeneity is worth exploring in greater detail.

Despite care in fitting the models, there are limitations to our analyses. With respect to our participant cohort, the iKM HER database we drew from – despite its wealth of information about community-based oncology – may be limited in various ways. For example, the iKM system is used for clinical practice purposes, not solely for research purposes. There may have been data entry errors at the point of care that could not be detected and hence corrected prior to analysis. Specific adverse events that may shed light on treatment responses are not graded in the iKM HER and oral therapies are recorded in or prescribed through iKM but the fulfillment of those prescriptions is not observable. In addition, we did not have overall survival data but rather only indices of death during the period of for which the participant was enrolled in the trial. In addition, patterns of care within the US Oncology Network, given its encouragement of evidence-based treatment guidelines, may differ to some degree from community-based treatment patterns in general.

The results of our studies suggest that there is real potential in leveraging routine clinical chemistries and demographic information in identifying individual participants that might be more likely to benefit, or not, from participation in a Phase I cancer clinical trial. However, there is clearly a need for replication and larger studies focusing on our results. In addition, although we considered routinely collected clinical chemistries and some demographic information in our analyses that are practical to measure, other factors, such as germline and tumor genome profiles, more comprehensive blood and plasma analyte collections, cell-free DNA (cfDNA), digital device data capturing mood, energy, pain, etc. should be considered to develop even more integrated and sophisticated models. We find evidence for differences in outcomes across the different trial indices but note that such an index would not be available for an individual seeking to enroll in a clinical trial of an intervention that has never been tested before. Obviously, greater insight into sources of heterogeneity among overall and individual outcomes across Phase I trials focusing on different interventions that is fueled by more in-depth profiling of the participants enrolled in them is needed as well as infrastructure to learn from experience in real time [11, 12]. Such a continuous ‘learning system’ would be of enormous benefit to clinical trialists and individual participants, and our studies provide motivation for seriously considering such an effort [44].

ACKNOWLEDGEMENTS

The authors would like to acknowledge internal support for this project from the Translational Genomics Research Institute (TGen).

FUNDING

No one had funding for this paper.

Chakiba, C., et al., Encouraging Trends in Modern Phase 1 Oncology Trials. N Engl J Med, 2018. 378(23): p. 2242-2243.
Glannon, W., Phase I oncology trials: why the therapeutic misconception will not go away. J Med Ethics, 2006. 32(5): p. 252-5.
Horstmann, E., et al., Risks and benefits of phase 1 oncology trials, 1991 through 2002. N Engl J Med, 2005. 352(9): p. 895-904.
Kurzrock, R. and R.S. Benjamin, Risks and benefits of phase 1 oncology trials, revisited. N Engl J Med, 2005. 352(9): p. 930-2.
Miller, F.G. and S. Joffe, Benefit in phase 1 oncology trials: therapeutic misconception or reasonable treatment option? Clin Trials, 2008. 5(6): p. 617-23.
Waligora, M., et al., Risk and surrogate benefit for pediatric Phase I trials in oncology: A systematic review with meta-analysis. PLoS Med, 2018. 15(2): p. e1002505.
Weber, J.S., et al., Reaffirming and Clarifying the American Society of Clinical Oncology's Policy Statement on the Critical Role of Phase I Trials in Cancer Research and Treatment. J Clin Oncol, 2017. 35(2): p. 139-140.
Tsimberidou, A.M., et al., Review of precision cancer medicine: Evolution of the treatment paradigm. Cancer Treat Rev, 2020. 86: p. 102019.
Chae, Y.K., et al., Path toward Precision Oncology: Review of Targeted Therapy Studies and Tools to Aid in Defining "Actionability" of a Molecular Lesion and Participant Management Support. Mol Cancer Ther, 2017. 16(12): p. 2645-2655.
Li, X. and J.L. Warner, A Review of Precision Oncology Knowledgebases for Determining the Clinical Actionability of Genetic Variants. Front Cell Dev Biol, 2020. 8: p. 48.
Schork, N.J., Artificial Intelligence and Personalized Medicine. Cancer Treat Res, 2019. 178: p. 265-283.
Schork, N.J., et al., Strategies for Testing Intervention Matching Schemes in Cancer. Clin Pharmacol Ther, 2020. 108(3): p. 542-552.
Iskander, R., et al., Survival Benefit Associated With Participation in Clinical Trials of Anticancer Drugs: A Systematic Review and Meta-Analysis. JAMA, 2024. 331(24): p. 2105-2113.
Shalowitz, D.I. and F.G. Miller, Are Participants With Cancer Best Managed in a Clinical Trial? JAMA, 2024. 331(24): p. 2077-2078.
Iskander, R., et al., The Benefits and Risks of Receiving Investigational Solid Tumor Drugs in Randomized Trials : A Systematic Review and Meta-analysis. Ann Intern Med, 2024. 177(6): p. 759-767.
Chau, N.G., et al., Early mortality and overall survival in oncology phase I trial participants: can we improve participant selection? BMC Cancer, 2011. 11: p. 426.
Bachelot, T., et al., Multivariable analysis of prognostic factors for toxicity and survival for participants enrolled in phase I clinical trials. Ann Oncol, 2000. 11(2): p. 151-6.
Bailey, C.H., et al., Progression-free Survival Decreases with Each Subsequent Therapy in Participants Presenting for Phase I Clinical Trials. J Cancer, 2012. 3: p. 7-13.
Brantner, C.L., et al., Comparison of methods that combine multiple randomized trials to estimate heterogeneous treatment effects. Stat Med, 2024.
Robertson, S.E., et al., Assessing Heterogeneity of Treatment Effects in Observational Studies. Am J Epidemiol, 2021. 190(6): p. 1088-1100.
Wendling, T., et al., Comparing methods for estimation of heterogeneous treatment effects using observational data from health care databases. Stat Med, 2018. 37(23): p. 3309-3324.
Oldenhuis, C.N., et al., Prognostic versus predictive value of biomarkers in oncology. Eur J Cancer, 2008. 44(7): p. 946-53.
Olmos, D., et al., Participant selection for oncology phase I trials: a multi-institutional study of prognostic factors. J Clin Oncol, 2012. 30(9): p. 996-1004.
Han, C., et al., Comparison of prognostic factors in participants in phase I trials of cytotoxic drugs vs new noncytotoxic agents. Br J Cancer, 2003. 89(7): p. 1166-71.
Stavraka, C., et al., Developing an objective marker to optimize participant selection and predict survival benefit in early-phase cancer trials. Cancer, 2014. 120(2): p. 262-70.
Grellety, T., et al., PRognostic factor of Early Death In phase II Trials or the end of 'sufficient life expectancy' as an inclusion criterion? (PREDIT model). BMC Cancer, 2016. 16(1): p. 768.
Wheler, J., et al., Survival of participants in a Phase 1 Clinic: the M. D. Anderson Cancer Center experience. Cancer, 2009. 115(5): p. 1091-9.
An, M.W. and S.J. Mandrekar, Time to progression ratio: promising new metric or just another metric? Ann Transl Med, 2016. 4(Suppl 1): p. S43.
Von Hoff, D.D., et al., Pilot study using molecular profiling of participants' tumors to find potential targets and select treatments for their refractory cancers. J Clin Oncol, 2010. 28(33): p. 4877-83.
Cirkel, G.A., et al., The time to progression ratio: a new individualized volumetric parameter for the early detection of clinical benefit of targeted therapies. Ann Oncol, 2016. 27(8): p. 1638-43.
Galecki, A. and T. Burzykowski, Linear Mixed-Effects Models Using R: A Step-by-Step Approach. 2015: Springer.
McCulloch, C.E., S.R. Searle, and J.M. Neuhuas, Generalized, Linear, and Mixed Models. 2008: Wiley.
van Buuren, S. and K. Groothuis-Oudshoorn, Multivariate Imputation by Chained Equations in R. Journal of Statistical Software, 2011. 45(3): p. 1-67.
Cox, D.R., Regression Models and Life-Tables Journal of the Royal Statistical Society, Series B., 1972. 34(2): p. 187-220.
Fischetti, T., E. Mayor, and R.M. Forte, R: Predictive Analysis. 2017: Packt Publishing.
Pedregosa , F., et al., Scikit-learn: Machine Learning in Python. JMLR, 2011. 12: p. 2825-2830.
Hastie, T., R. Tibshirani, and J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed. 2008: Springer.
van der Laan, M.J., E.C. Polley, and A.E. Hubbard, Super learner. Stat Appl Genet Mol Biol, 2007. 6: p. Article25.
Wolpert, D.H., Stacked Generalization. Neural Networks, 1992. 5(2): p. 241-259.
Dablain, D., B. Krawczyk, and N.V. Chawla, DeepSMOTE: Fusing Deep Learning and SMOTE for Imbalanced Data. IEEE Trans Neural Netw Learn Syst, 2022. PP.
Chawla, N.V., et al., SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence, 2002. 16: p. 321-357.
Fernandez, A., et al., SMOTE for Learning from Imbalanced Data: Progress an dChallenges, Marking the 15-year Anniversary. Journal of Artificial Intelligence Research 2018. 61: p. 863-905.
Pepe, M.S., The Statistical Evaluation of Medical Tests for Classification and Prediction. 2004: Oxford University Press.
Shrager, J. and J.M. Tenenbaum, Rapid learning for precision oncology. Nat Rev Clin Oncol, 2014. 11(2): p. 109-18.
Cohen, J., A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 1960. 20: p. 37-46.

Competing interest reported. [email protected], [email protected], [email protected], [email protected], [email protected], are employees of McKesson/Ontada from which the data were obtained.

SUPPLEMENTARYMETHODS.docx

Predicting Individual Responses in Phase I Oncology Trials Using Routinely Collected Clinical Biomarkers

Status:

Version 1

Abstract

Figures

INTRODUCTION

METHODS

Participant Data Sources

Responses and Candidate Prognostic Biomarkers

Analytical Methods

RESULTS

Final Data Set

Initial Trial Index and Tumor Type Analysis

Machine Learning Model Analysis

Mixed Model Analysis

Sensitivity Analysis

Example Predictions

DISCUSSION

Summary of Findings

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1