survivalsm function fails using another dataset

[ShengnanHuang9]

Hi @mnwright,

Thanks for sharing this package and I've been trying to use it on my own dataset. However, when using the code like this:
survsvm.reg1 <- survivalsvm(Surv(follow_up, event) ~ AGE.bl+PTGENDER+PTEDUCAT , data = study1, type = "regression", gamma.mu = 1, opt.meth = "quadprog", kernel = "add_kernel")
An error shows up:

There are definitely observations in the dataset study1, and follow_up and event are all num.

Could you please let me know where goes wrong? Thanks!

[fouodo]

Hi, I tried it myself with this example using the following code: survivalsvm(Surv(time, status) ~ karno + age, veteran, gamma.mu = 0.1) and obtained this output: survivalsvm result Call: survivalsvm(Surv(time, status) ~ karno + age, veteran, gamma.mu = 0.1) Survival svm approach : regression Type of Kernel : lin_kernel Optimization solver used : quadprog Number of support vectors retained : 15 survivalsvm version : 0.0.5 So it is not easier to figure out what could the raison in your case. May be it would be nice to send us more details about it. Regards Cesaire

…

On 2. Aug 2019, at 23:04, certara-ShengnanHuang ***@***.***> wrote: Hi @mnwright <https://github.com/mnwright>, Thanks for sharing this package and I've been trying to use it on my own dataset. However, when using the code like this: survsvm.reg1 <- survivalsvm(Surv(follow_up, event) ~ AGE.bl+PTGENDER+PTEDUCAT , data = study1, type = "regression", gamma.mu = 1, opt.meth = "quadprog", kernel = "add_kernel") Error shows up a this: <https://user-images.githubusercontent.com/50929185/62398613-3780bc80-b547-11e9-8708-f96248cef140.png> There are definitely observations in the dataset study1, and follow_up and event are all num. Could you please let me know where goes wrong? Thanks! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#6?email_source=notifications&email_token=AGJ42MHJV4DYZHOD4MKXGATQCSOMNA5CNFSM4IJBA4T2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HDETOHA>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGJ42MCSOKTM7M3VVP6KQMDQCSOMNANCNFSM4IJBA4TQ>.

[ShengnanHuang9]

Hi Cesaire,

Thanks for your reply. I think I found the reason why. The function only accepts the covariate as a factor or continuous variable.
I used this function and get the predicted out as the predicted event time (not the censoring time). Have you ever tried to plot the predicted KM vs. the raw KM on the dataset? I try using this formula:

survsvm.reg <- survivalsvm(Surv(diagtime, status) ~ .,
                           data = veteran,
                           type = "regression", gamma.mu = 1,
                           opt.meth = "quadprog", kernel = "add_kernel")
survsvm.ptrain <- predict(object = survsvm.reg, newdata = veteran)

And then plot the predicted and the raw KM for veteran dataset as follow, you could see they are significantly different:

I also calculated the Brier score, only 0.04 - a really good sign for model performance, which contradicts a little with the KM curve. Is this result reasonable for your package?

Besides is the survsvm.ptrain$predicted only the observed event time or it is the observed event or censoring time, corresponding to the status indicator? Or it is just the risk level, not the predicted time at all?

Thanks in advance!

Regards,
Shengnan

[fouodo]

HI Certara, Survival SVM does not predict harzard or any survival probability, but survival ranks. So C-Index as presented in our paper (Support Vector Machines for Survival Analysis with R - The R Journalhttps://journal.r-project.org/archive/2018/RJ.../RJ-2018-005.pdf <https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=1&cad=rja&uact=8&ved=2ahUKEwj4oIXK0-3jAhVBhqQKHYMHB8oQFjAAegQIBRAC&url=https%3A%2F%2Fjournal.r-project.org%2Farchive%2F2018%2FRJ-2018-005%2FRJ-2018-005.pdf&usg=AOvVaw3KelwnvXdZ920Db6Tm5Rfx>) should be the more appropriate performance measure between survival sim and others model like Cox or random survival forest. Best regards Cesaire

…

On 5. Aug 2019, at 21:02, certara-ShengnanHuang ***@***.***> wrote: Hi Cesaire, Thanks for your reply. I think I found the reason why. The function only accepts covariates as a factor or continuous variable. I used this function and get the predicted out as the predicted event time. Have you ever tried to plot the predicted KM vs. the raw KM on the dataset? I try using this formula: survsvm.reg <- survivalsvm(Surv(diagtime, status) ~ ., data = veteran, type = "regression", gamma.mu = 1, opt.meth = "quadprog", kernel = "add_kernel") survsvm.ptrain <- predict(object = survsvm.reg, newdata = veteran) And then plot the predicted and the raw KM for veteran dataset as follow, you could see they are significantly different: <https://user-images.githubusercontent.com/50929185/62487927-13b0b700-b791-11e9-86e6-3d3dce72d360.png> I also calculated the Brier score, only 0.04 - a really good sign for model performance, which contradicts a little with the KM curve. Is this result reasonable for your package? Regards, Shengnan — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#6?email_source=notifications&email_token=AGJ42MDDZMTRAMPOKUAIHDDQDB2L7A5CNFSM4IJBA4T2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD3SYJLY#issuecomment-518358191>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGJ42MHHG56SUX52QSNTR6LQDB2L7ANCNFSM4IJBA4TQ>.

[ShengnanHuang9]

Hi Cesaire,

Thanks for the explanation, it really makes sense now.
C-index is really useful and the Brier score is also popular for performance measure between different models. Is there a way to calculate the Brier score for your survival SVM?
Thanks a lot!

Best regards,
Shengnan

[fouodo]

Hi Certara, Unfortunately no. It doesn’t make sense to compute brier score on survivalsvm output. Best Cesaire

…

On 6. Aug 2019, at 15:24, certara-ShengnanHuang ***@***.***> wrote: Hi Cesaire, Thanks for the explanation, it really makes sense now. C-index is really useful and the Brier score is also popular for performance measure between different models. Is there a way to calculate the Brier score for your survival SVM? Thanks a lot! Best regards, Shengnan — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#6?email_source=notifications&email_token=AGJ42MBJET6FXGVVCPKF3KTQDF3RZA5CNFSM4IJBA4T2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD3VD4ZY#issuecomment-518667879>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGJ42MG2D27GQKLH36D2XVDQDF3RZANCNFSM4IJBA4TQ>.

[ShengnanHuang9]

Hi Cesaire,

That is a pity... Thanks for your reply.

Best,
Shengnan