For a study on CSF frontotemporal dementia (FTD) biomarkers, they managed to collect a relatively fair number of patients. Also, the postmortem confirmation is a strong point, as FTD is heterogeneous and clinical diagnosis is not simple. However, for the kind of conclusions (very high sensitivity and specificity), the study is not powered enough (although we realize it is not easy to obtain high numbers of characterized, postmortem confirmed FTD patients). In the end, the sensitivity and specificity numbers are based on just 10 FTD patients versus 10 AD patients, and thus every single patient may influence these numbers strongly.

Their approach to use the natural log to transform ELISA data to Luminex data leads to a good fit. But, on the other hand, there is an overfitting of the data: In the majority of figures of the Ln-transformed data of the test sets in Figure 2, the correlation coefficient is high due to a single outlier, and the trend lines are thus too steep. Since the correlation coefficients of the test...  Read more

For a study on CSF frontotemporal dementia (FTD) biomarkers, they managed to collect a relatively fair number of patients. Also, the postmortem confirmation is a strong point, as FTD is heterogeneous and clinical diagnosis is not simple. However, for the kind of conclusions (very high sensitivity and specificity), the study is not powered enough (although we realize it is not easy to obtain high numbers of characterized, postmortem confirmed FTD patients). In the end, the sensitivity and specificity numbers are based on just 10 FTD patients versus 10 AD patients, and thus every single patient may influence these numbers strongly.

Their approach to use the natural log to transform ELISA data to Luminex data leads to a good fit. But, on the other hand, there is an overfitting of the data: In the majority of figures of the Ln-transformed data of the test sets in Figure 2, the correlation coefficient is high due to a single outlier, and the trend lines are thus too steep. Since the correlation coefficients of the test sets are therefore much lower than the training set, it is clear that the current transformation model cannot be extrapolated to other cohorts or even single patients. This is clearly also due to the fact that the training set of 40 data points does not provide sufficient power to obtain a generalizable formula.

Finally, we find that the aim not very clearly stated: Is it to compare two platforms? Or is it to define cutoff points for discriminating AD from FTD? The title may be taken as the aim is a comparison of the two platforms. Instead, what they did is a merging of data obtained from two different platforms, which are not easily interchangeable (Jongbloed et al., in press).

References:
Jongbloed W et al. Discriminatory and Predictive Capabilities of ELISA and Multiplex Platforms in a Longitudinal Alzheimer Study. Alzheim & Dement. 2012, in press.

View all comments by Wesley Jongbloed
View all comments by Philip Scheltens
View all comments by Charlotte Teunissen