A Prognostic Model for Overall Survival in Men with Metastatic Castration-resistant Prostate Cancer - Susan Halabi
August 11, 2019
Susan Halabi provides an update on an externally validated prognostic model of overall survival (OS) in men with metastatic castration-resistant prostate cancer (mCRPC) treated with docetaxel, which included eight important prognostic factors variables of overall survival: opioid analgesic use, ECOG performance status, albumin, disease site, LDH, hemoglobin, PSA, and alkaline phosphatase. They used this model to develop prognostic risk groups and sought to externally validate this model in a broader group of men with mCRPC, in specific subgroups (White, Black, Asian patients, different age groups), and to validate the two and three prognostic risk groups in this large dataset.
Biographies:
Susan Halabi, PhD, Professor of Biostatistics and Bioinformatics, Duke Cancer Institute, Durham, North Carolina, United States
Charles J. Ryan, MD, the President and Chief Executive Officer of The Prostate Cancer Foundation (PCF), the world’s leading philanthropic organization dedicated to funding life-saving prostate cancer research. Charles J. Ryan is an internationally recognized genitourinary (GU) oncologist with expertise in the biology and treatment of advanced prostate cancer. Dr. Ryan joined the PCF from the University of Minnesota, Minneapolis, where he served as Director of the Hematology, Oncology, and Transplantation Division in the Department of Medicine. He also served as Associate Director for Clinical Research in the Masonic Cancer Center and held the B.J. Kennedy Chair in Clinical Medical Oncology.
Biographies:
Susan Halabi, PhD, Professor of Biostatistics and Bioinformatics, Duke Cancer Institute, Durham, North Carolina, United States
Charles J. Ryan, MD, the President and Chief Executive Officer of The Prostate Cancer Foundation (PCF), the world’s leading philanthropic organization dedicated to funding life-saving prostate cancer research. Charles J. Ryan is an internationally recognized genitourinary (GU) oncologist with expertise in the biology and treatment of advanced prostate cancer. Dr. Ryan joined the PCF from the University of Minnesota, Minneapolis, where he served as Director of the Hematology, Oncology, and Transplantation Division in the Department of Medicine. He also served as Associate Director for Clinical Research in the Masonic Cancer Center and held the B.J. Kennedy Chair in Clinical Medical Oncology.
Read the Full Video Transcript
Charles Ryan: Hello from ASCO 2019. I'm delighted to be joined by Dr. Susan Halabi from Duke University, a colleague and friend, we've worked together for a long time. You've done some really important work for the field on a number of different levels, but I think one of the things that the clinicians know you best for is the Nomogram, the model that you developed, which is really a multi-variable model for outcome for patients that allows us to look at factors in the blood, PSA levels, alkaline phosphatase, etc., and determine how important they are in terms of the patient's prognosis. I actually use this model almost every time I give a lecture and I understand you've updated it recently. Tell us about your updates.
Susan Halabi: Thank you, Dr. Ryan, for the introduction. I'm also delighted to give you an update on where we are. Actually, we haven't yet updated the model. What we have previously, as you've mentioned, we have developed a prognostic model of overall survival, and that model included eight variables that are important prognostic factors of overall survival. The model was developed using data from CALGB 90401 and it was externally validated using the INFUSE 33 trial. So what, we set to do here is to validate the model using more data sets. So we use additional four data sets from different trials, they're all randomized phase three trials that included, SYNERGY, Mainsail, the READY trial, and we also brought in the fourth data set, which is a trial from the Tasquinimod trial. So this is the first data set that includes data from a non-docetaxel treatment.
Charles Ryan: So you're really trying to find data from a number of trials from which the treatments are different because you're looking for prognostic variables that are important for survival, not necessarily predictive variables that would help a clinician determine the efficacy of a treatment. I just want to kind of make that point.
Susan Halabi: Yes, that is correct. The model was developed to predict overall survival, but these variables are prognostic. So it's independent of treatment. However, in this updated analysis that's going to be presented at ASCO tomorrow, this is abstract 5022, what we've done is we looked at the performance of the model. We measured the area under the curve in around 5500 patients in total, and then we looked, for the first time we've looked at the performance of the model in subgroups of patients, by race and by age group and by treatment type. And overall what we found out is the model is very well robust. We have an area under the curve of about ... ranging anywhere from .75 to .79, and the model is definitely well-calibrated. However, as you know, the variables included in this model are mostly clinical variables. So definitely there is a need to include biomarker data and hopefully, this will be done in the future.
Charles Ryan: Well, we have biomarkers in the model. They're just, we don't think of them as biomarkers because they're standard blood tests, which I find interesting. LDH, alkaline phosphatase, hemoglobin, all very, very important biomarkers, but they're not new biomarkers really.
Susan Halabi: That is correct. So these are what we as statisticians refer to as low dimensional variables. We would like to use more molecular data from high dimensional data.
Charles Ryan: What can you tell us about what you found as you looked across different racial groups, different geographic groups. You said that the model's very robust. I think by that you're indicating that a clinician could use your model if he's treating an African American patient, he doesn't have to worry that the model is predicting only outcomes in Caucasians or something like that, right?
Susan Halabi: That is correct. So the advantage of doing this big analysis based on about 5500 patients is that we have a high number of African American man, which was unlike how the model was developed in the original trial. In the original trial, we had about 10% of the patients were African American. So when you look at the model, the area under the curve and overall in Asian and African American men and in Caucasian man were all about .75, which is very similar to the overall performance that we observed in the whole trial, in the original trial. So it did perform very well across these racial groups. And also when we looked at the model performance by age group, we also found that regardless of the age of the patient, the model was very robust.
Charles Ryan: Very important.
Susan Halabi: We also looked at the docetaxel versus non-docetaxel trials, and again, the model did very well in docetaxal-only patients. It was maybe a little bit higher but really nothing of significance, just nominally higher. And in the Tasquinimod trial, it was also about .75. So we're right around the average that we found in the original data.
We also had the opportunity to validate also the two and the three prognostic risk factors. So originally when we developed the model, we also came up with prognostic groups. So we validated the low and the two and the three risk groups. The two risk groups will classify patients into low and high risk groups, whereas the three risk group would classify the patient into low, intermediate, and high. And we really saw a very nice separation across all the trials. You can look at the Kaplan Meier curves and they're more widely separated, which attest to the performance of the model.
Charles Ryan: So should we be doing that in clinical trials moving forward for CRPC? Should we stratify by low, intermediate, or high risk when we develop a trial?
Susan Halabi: Well actually, that's a very good question, and I think it is definitely recommended to do that. A lot of the trials we have used this model, for instance for the A031201, we have used the prognostic groups as a stratification value in the randomization. And when you do that, what you're doing is you're assuring that the distribution of the patients across the risk groups is balanced between the arms so you don't have to worry about imbalances and prognostic factors.
I guess the main advantage, and this is often a question raised by clinicians, why would I use a model as opposed to using one variable. Well, the main advantage of using the model, you're factoring all these variables together so you don't have to worry to bring in eight certification variables in your randomization, which will not be feasible. Whereas when you look at risk group, you're only introducing one variable in your stratification.
Charles Ryan: Well, as I say, when I give lectures in the community or to trainees, I show your model and I have a slide that says what's the prognosis. And I show the variability of the prognosis for a patient in CRPC because I think a lot of clinicians underappreciate the spectrum of prognosis that you've shown that does exist in CRPC. And while clinical trial development is one thing, for the clinicians who are watching this, my sort of take-home lesson is that you should be monitoring these variables. That's the point. I think a lot of the physicians may not routinely monitor them because I say, if a patient's PSA is going up a little bit but their alkaline phosphatase is going up, those things work in concert to make the prognostic model more adverse for that patient. So it's an open question as to how this should be used in a daily clinical setting, but I think we should be collecting that data and putting that into our sort of mental framework as we make decisions for the patient.
Susan Halabi: I totally agree with you. I think one ... another maybe suggestion for how people may use it is how you want to treat your patients. So if you run the model at baseline and you find that the patient is high risk, you may need to treat them more aggressively than someone who is low risk. But again, I think even though this model is very robust and doing well, I think there is definitely the need to do better than the .75. It will have been nice to develop a new model where we can eliminate the variability and bring in the area under the curve up to about maybe higher than .80, maybe 0.85 or higher.
Charles Ryan: That'll be good for us in terms of counseling our patients, it'll be good for the patients, and it'll be good for designing clinical trials that really get answers faster and more accurately. So I want to thank you for that work. Thank you for sitting down with me. It's always a pleasure. And congratulations on getting this moving forward.
Susan Halabi: Thank you very much. It's my pleasure, Dr. Ryan, as always. Thank you.
Charles Ryan: Hello from ASCO 2019. I'm delighted to be joined by Dr. Susan Halabi from Duke University, a colleague and friend, we've worked together for a long time. You've done some really important work for the field on a number of different levels, but I think one of the things that the clinicians know you best for is the Nomogram, the model that you developed, which is really a multi-variable model for outcome for patients that allows us to look at factors in the blood, PSA levels, alkaline phosphatase, etc., and determine how important they are in terms of the patient's prognosis. I actually use this model almost every time I give a lecture and I understand you've updated it recently. Tell us about your updates.
Susan Halabi: Thank you, Dr. Ryan, for the introduction. I'm also delighted to give you an update on where we are. Actually, we haven't yet updated the model. What we have previously, as you've mentioned, we have developed a prognostic model of overall survival, and that model included eight variables that are important prognostic factors of overall survival. The model was developed using data from CALGB 90401 and it was externally validated using the INFUSE 33 trial. So what, we set to do here is to validate the model using more data sets. So we use additional four data sets from different trials, they're all randomized phase three trials that included, SYNERGY, Mainsail, the READY trial, and we also brought in the fourth data set, which is a trial from the Tasquinimod trial. So this is the first data set that includes data from a non-docetaxel treatment.
Charles Ryan: So you're really trying to find data from a number of trials from which the treatments are different because you're looking for prognostic variables that are important for survival, not necessarily predictive variables that would help a clinician determine the efficacy of a treatment. I just want to kind of make that point.
Susan Halabi: Yes, that is correct. The model was developed to predict overall survival, but these variables are prognostic. So it's independent of treatment. However, in this updated analysis that's going to be presented at ASCO tomorrow, this is abstract 5022, what we've done is we looked at the performance of the model. We measured the area under the curve in around 5500 patients in total, and then we looked, for the first time we've looked at the performance of the model in subgroups of patients, by race and by age group and by treatment type. And overall what we found out is the model is very well robust. We have an area under the curve of about ... ranging anywhere from .75 to .79, and the model is definitely well-calibrated. However, as you know, the variables included in this model are mostly clinical variables. So definitely there is a need to include biomarker data and hopefully, this will be done in the future.
Charles Ryan: Well, we have biomarkers in the model. They're just, we don't think of them as biomarkers because they're standard blood tests, which I find interesting. LDH, alkaline phosphatase, hemoglobin, all very, very important biomarkers, but they're not new biomarkers really.
Susan Halabi: That is correct. So these are what we as statisticians refer to as low dimensional variables. We would like to use more molecular data from high dimensional data.
Charles Ryan: What can you tell us about what you found as you looked across different racial groups, different geographic groups. You said that the model's very robust. I think by that you're indicating that a clinician could use your model if he's treating an African American patient, he doesn't have to worry that the model is predicting only outcomes in Caucasians or something like that, right?
Susan Halabi: That is correct. So the advantage of doing this big analysis based on about 5500 patients is that we have a high number of African American man, which was unlike how the model was developed in the original trial. In the original trial, we had about 10% of the patients were African American. So when you look at the model, the area under the curve and overall in Asian and African American men and in Caucasian man were all about .75, which is very similar to the overall performance that we observed in the whole trial, in the original trial. So it did perform very well across these racial groups. And also when we looked at the model performance by age group, we also found that regardless of the age of the patient, the model was very robust.
Charles Ryan: Very important.
Susan Halabi: We also looked at the docetaxel versus non-docetaxel trials, and again, the model did very well in docetaxal-only patients. It was maybe a little bit higher but really nothing of significance, just nominally higher. And in the Tasquinimod trial, it was also about .75. So we're right around the average that we found in the original data.
We also had the opportunity to validate also the two and the three prognostic risk factors. So originally when we developed the model, we also came up with prognostic groups. So we validated the low and the two and the three risk groups. The two risk groups will classify patients into low and high risk groups, whereas the three risk group would classify the patient into low, intermediate, and high. And we really saw a very nice separation across all the trials. You can look at the Kaplan Meier curves and they're more widely separated, which attest to the performance of the model.
Charles Ryan: So should we be doing that in clinical trials moving forward for CRPC? Should we stratify by low, intermediate, or high risk when we develop a trial?
Susan Halabi: Well actually, that's a very good question, and I think it is definitely recommended to do that. A lot of the trials we have used this model, for instance for the A031201, we have used the prognostic groups as a stratification value in the randomization. And when you do that, what you're doing is you're assuring that the distribution of the patients across the risk groups is balanced between the arms so you don't have to worry about imbalances and prognostic factors.
I guess the main advantage, and this is often a question raised by clinicians, why would I use a model as opposed to using one variable. Well, the main advantage of using the model, you're factoring all these variables together so you don't have to worry to bring in eight certification variables in your randomization, which will not be feasible. Whereas when you look at risk group, you're only introducing one variable in your stratification.
Charles Ryan: Well, as I say, when I give lectures in the community or to trainees, I show your model and I have a slide that says what's the prognosis. And I show the variability of the prognosis for a patient in CRPC because I think a lot of clinicians underappreciate the spectrum of prognosis that you've shown that does exist in CRPC. And while clinical trial development is one thing, for the clinicians who are watching this, my sort of take-home lesson is that you should be monitoring these variables. That's the point. I think a lot of the physicians may not routinely monitor them because I say, if a patient's PSA is going up a little bit but their alkaline phosphatase is going up, those things work in concert to make the prognostic model more adverse for that patient. So it's an open question as to how this should be used in a daily clinical setting, but I think we should be collecting that data and putting that into our sort of mental framework as we make decisions for the patient.
Susan Halabi: I totally agree with you. I think one ... another maybe suggestion for how people may use it is how you want to treat your patients. So if you run the model at baseline and you find that the patient is high risk, you may need to treat them more aggressively than someone who is low risk. But again, I think even though this model is very robust and doing well, I think there is definitely the need to do better than the .75. It will have been nice to develop a new model where we can eliminate the variability and bring in the area under the curve up to about maybe higher than .80, maybe 0.85 or higher.
Charles Ryan: That'll be good for us in terms of counseling our patients, it'll be good for the patients, and it'll be good for designing clinical trials that really get answers faster and more accurately. So I want to thank you for that work. Thank you for sitting down with me. It's always a pleasure. And congratulations on getting this moving forward.
Susan Halabi: Thank you very much. It's my pleasure, Dr. Ryan, as always. Thank you.