Answers are grounded in this paper and cite their sources.
Magnetic resonance imaging (MRI)-based hippocampus volume (HV) is the best established imaging marker to support the prediction of AD dementia (ADD) in mild cognitive impairment (MCI), although its utility in clinical patient care has not yet been fully demonstrated [1], [2]. HV can be scored on an ordinal scale based on visual inspection of MRI [3], or it can be estimated quantitatively by manual or automatic delineation of the hippocampus in MRI. While visual scoring tends to perform worse in MCI-to-ADD prediction, manual delineation and automatic methods show very similar performance [4]. Furthermore, there is hardly any difference among the numerous automatic methods with respect to predictive power in MCI. In the head-to-head comparison of four HV measurement methods in MCI subjects of the Alzheimer's Disease Neuroimaging Initiative by the European Medicines Agency, the area (AUC) under the receiver operating characteristic (ROC) curve for 2-year prediction of ADD ranged between 0.7290 and 0.7565, and among three of the four methods, the AUC ranged between 0.7516 and 0.7565 [5]. This appears surprising at first sight given that the quantitative methods differ strongly in accuracy and precision with respect to the anatomical delineation of the hippocampus. Time spent by the rater and computer processing time also differ strongly [4]. Here, we aim to provide a simple mathematical explanation of the stability of the performance of MRI-based HV in MCI with respect to the HV measurement method. The AUC of MCI-to-ADD prediction by the true HV is 0.7602 in this scenario (Fig. 1A), close to the European Medicines Agency results cited previously. Part (A) shows the ROC curve for MCI-to-ADD prediction by the true, error-free HV assumed to follow Gaussian distributions according to (4). Part (B) illustrates the deterioration of predictive performance (as measured by the decline of the area AUC under the ROC curve) by additional intersubject variability due to measurement errors σmeas of HV estimates. The red curve shows the impact of the additional variability on the performance of the true HV according to part (A). The other curves show the impact of measurement error when the true, error-free HV would perform better (actual biological group difference Dtrue > 1) or worse (Dtrue < 1) than assumed in part (A). Abbreviations: HV, hippocampus volume; ADD, AD dementia; AUC, area under the ROC curve; MCI, mild cognitive imapairment. The red curve in Fig. 1B shows the AUC for prediction of MCI-to-ADD progression by HV estimates as function of σmeas (scaled to μS − μP). The additional variability by the measurement error causes only mild AUC decrease. Even when the standard deviation of the measurement error reaches the difference μS − μP of mean true HV between MCI stables and MCI-to-ADD progressors, the AUC only slightly decreases to 0.6915 (from AUC = 0.7602 with true, error-free HV), which still lies within the typically observed performance range. To assess the impact of additional variance on the prognostic performance also for other biological scenarios, σtrue was varied whereas μS and μP were kept constant at the values in (4). Biological scenarios were labeled by the actual biological group difference scaled to the actual (patho)physiological variance Dtrue = (μS − μP)/σtrue (Fig. 1B). It is evident from Fig. 1B that the impact of the additional variance caused by the measurement process decreases with decreasing biological group difference. The lower the initial AUC, the flatter the decline of the AUC with increasing measurement error. This suggests that the stability of the predictive power of MRI-based HV in MCI with respect to the measurement method is a floor effect: even with the best measurement method, the power of MRI-based HV to predict MCI-to-ADD progression is inherently limited by the predictive properties of hippocampal atrophy. Thus, further decreasing HV measurement error compared with existing methods will have only very little impact on the predictive accuracy of hippocampus volumetry. To make HV widely available for routine clinical use, the measurement method should combine ease of use and short computation time with acceptable accuracy and precision. Efforts to harmonize HV measurement in the context of AD might account for this [7]. Furthermore, integrating existing HV volumetry methods in multivariable models rather than increasing HV measurement accuracy will be most efficient to make the best use of MRI-based HV as prognostic marker in MCI. Finally, hippocampal atrophy is not homogeneous across hippocampal subfields, suggesting that MRI-based volume measures of specific hippocampal subfields might provide better predictive power compared with the entire hippocampus [8].
No takes yet. Share an insight, caveat, or question.
Buchert et al. (2018) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: