© 2004 American Thoracic Society
Prognosis in Idiopathic Pulmonary FibrosisTo the Editor:In a recent study, Collard and colleagues (1) described new prognostic tools for patients with idiopathic pulmonary fibrosis. Although this study may provide new insights into the course of this disease, we feel that such scoring systems are of doubtful value in clinical practice. Concerns raised by this study relate to statistical limitations of prognostic models and clinical utility of these tools. Statistical pitfalls of prognostic models have been extensively reviewed and account for the disappointing results of such models when applied to patients other than those used to build the model (2, 3). For instance, in the study by Collard and colleagues (1), many variables have been tested for association with survival, leading to a considerable inflation of the type I error rate. In other words, chance alone may explain the association between survival and several of the tested variables. In addition, in its simplest form, the Cox model assumes that the logarithm of the hazard function is linearly related to each predictor. This implies that a decrease of FVC from 100 to 90% carries the same prognostic information as a decrease from 50 to 40%. Such an assumption seems unrealistic on clinical grounds. In connection with these limitations and uncertainties, whether these results are transposable to similar patients in another location (4) remains an important question. The second concern relates to clinical application of these results, even if the model is valid. For example, the model tells us that 46.4% of patients with a 10% 6-month decrease in FVC will survive at 5 years. This means that, among a large group of similar patients, about 46% will survive 5 years (with a great uncertainty given the wide confidence interval). For an individual patient, it tells us that there is a 46% chance of surviving 5 years. This does not provide much certainty about whether the patient will survive that long. In conclusion, Collard and colleagues (1) have found several variables to be associated with survival in their cohort of patients. Whether these results could be replicated on other sets of patients remains to be demonstrated. The use of these tools at bedside is another story.
Hôpital Beaujon Clichy, France FOOTNOTES Conflict of Interest Statement: G.T. and M.F. have no declared conflict of interest. REFERENCES
From the Authors: We appreciate the careful reading Drs. Thabut and Fournier have given our recent publication (1). Their concerns about the statistical limitations and clinical utility of our findings are important and deserve a response. Statistical analysis of multiple variables may lead to an increased chance of incorrectly identifying predictors (type I error). To limit the impact of this, as we described in the METHODS, we initially estimated the hazard ratios for each potential predictor variable. To maintain adequate power, this hazard ratio was then adjusted for only baseline value, age, and smoking sttus, thereby limiting potential inflation of the type I error rate. Model fit comparisons were then performed to identify the variables that best predicted survival. Even had we corrected for multiple comparisons using the false discovery rate procedure (2), only the 6-month change in PaO2 would have been a potentially false discovery. Thus, it is highly unlikely that chance alone played a role in our findings. However, given the limitations and uncertainties inherent in these types of studies, we agree that it is important to confirm the findings in other larger populations. Importantly, several recent studies, two of which were published in the same issue of the Journal as our study, support our findings (3–6). The ability of predictors of survival to prognosticate in individual patients is, of course, limited. Given the complex nature of disease and the myriad factors impacting survival, no individual or group of predictors will tell individual patients exactly how long they have to live. However, if one steps away from making exact predictions, surely Drs. Thabut and Fournier would agree that it is clinically useful to know that worsening dyspnea, a decreasing FVC, or an increasing A-a gradient carry a worse prognosis than stable or improving values over standard lengths of time. In the absence of a perfect prognostic tool, we consider these data to be an advance in the management of patients with idiopathic pulmonary fibrosis and clinically useful in counseling of patients regarding prognostic expectations, providing guidance regarding the usefulness of a therapeutic approach, and determining the timing of transplantation listing.
University of Colorado Health Sciences Center Denver, Colorado FOOTNOTES Conflict of Interest Statement: H.R.C. and K.K.B. have no declared conflict of interest. REFERENCES
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
