Im involved in a project which we desired to answer the follwoing: what is the probability of this patient to be sucessfully free of ventilation support 2 days from now, or 4 days from now or a week from now, if the ventilation support team starts its removal right now?
I have data from a cohort where many predictors, such as ventilator sets, blood gases etc, are measured every two days up to ventilation support successful removal. Therefore, the dataset has a typical predictors time dependent structure.
The first question is: is it feasible to use such data to develop a prediction model? I looked around and found some stuff supporting in the positive answer.
http://www.ncbi.nlm.nih.gov/pubmed/15813715
http://arxiv.org/abs/1306.6479
But I also found some stuff supporting a negative answer:
http://www.annualreviews.org/doi/abs/10.1146/annurev.publhealth.20.1.145?journalCode=publhealth
So far, a believe that it is possible, and in such a model the outcome prediction can be made at any time, not only at baseline as in traditional right censored survival model. Am I correct? However, Im not completely sure that cox proportional hazard model is adequate for this purpose. If it is not, what are the alternatives?
But then other questions follow.
To do time dependent predictions, I guess I must set the baseline date of every single patient to 1, and the following dates where predictors are repeatedly measured must be set as differences to his/her baseline date, in a way that, similar to the traditional right censored model, all patients will start at day 1. Is this reasonable?
The last question. Suppose I have developed and validated this model, and I have five different risk groups as in this paper
http://jco.ascopubs.org/content/25/23/3503.full.pdf
If the patient is classified at group low risk, but after a while in this group without the outcome, he/she changes the predictor values and this moves he/she to another risk group. Does the risk estimation for her/him will be like just jumping from one survival curve to another at the same time? Or his/her time will be set to baseline time in the survival curve as he/she are just starting the follow up right now?
Another way to ask the same thing...the patient is at baseline and I want to estimate the probability of the outcome at day 2 from now. This is simple. Now suppose that he/she did not have the outcome and up to day 2 of the follow-up. Now I want to predict the outcome in the next two days with different predictors values. Should I predict for 2 days or should predict the outcome at day 4 of the follow-up?
I was challenged with this question. For me, so far, the correct way is to predict at day 4 (or just jump from one survival curve to another at the same time), as it correspond to "calendar" date of that patient in the follow-up, as it is in the dataset. However, I heard an argument that made me wonder. This argument is that survival models deal with conditional probabilities in a way that if this patient did not have the outcome up to now, the probability for the next two days are conditioned to his survival. Therefore, predictions should be made for his next two days, not at day 4.
Any light?
Dear Pedro,
Of course previous failure to wean a patient from the ventilator affects the survival predictions but there are exceptions ( see my paper: Miss B and Miss P). You may benefit from reading these papers:
Article Noninvasive Ventilation for Persistent Weaning Failure
Article Noninvasive Ventilation during Persistent Weaning Failure
Book Noninvasive Mechanical Ventilation and Difficult Weaning in ...
Article Efficacy of two noninvasive weaning strategies in intubated ...
Article The differential diagnosis for failure to wean from mechanic...
Dear beatrice,
Thank you for your feedback. I did read some stuff you posted, however, at the moment, my concern are the survival or other dynamic model suiting to answer my problem of individual predictions and not practice on ventilation support practice. The stuff I read, did not give me a different light so far. It turns that after a while banging around, I found some answers I was looking for in Therneau's book on survival analysis. Yes it is possible to make individual predictions with time dependent predictors. To set or not to set the baselines dates of each patient is a choice whether to consider a calendar date or to assume that there is no cohort effect and all patient start counting the time from a certain event such as hospital admission or aptitude of ventilatory support removal. I my case It would be more interesting set a baseline for everyone instead of using a calendar date. But there is some particular stuff to consider. At last, the individual predictions will always count from time zero of each patient. This means that, if the analysis considers risk groups, and the patient, at some time at the follow-up, has a change in the predictor value, he jumps from one risk group to another at that particular time. Other issues rise, for example, suppose I am at baseline and I want to make individual predictions at day 2, 4 and 7. I will have to guess reasonable values for the time dependent predictors at the days 2 and 4 (assuming I have the values from the baseline) . At first, carry over may be reasonable strategy, but updating the prediction will be required as soon as the information changes as it may substantially change the following survival probabilities. Another issue, is that Im not sure so far what calibration and performance measures are up to the validation phase of such a time dependent model, as, besides R2, all other measures that I found only works for the traditional right censored data. I will eventually read more of the stuff you sent to become more aware of others think and conduct analysis on this subject. Regards.
- Badges
- Science topic
- Similar topics
- Correction