Bias-adjustment in neuroimaging-based brain age frameworks: A robust scheme

Abstract

The level of prediction error in the brain age estimation frameworks is associated with the authenticity of statistical inference on the basis of regression models. In this paper, we present an efficacious and plain bias-adjustment scheme using chronological age as a covariate through the training set for downgrading the prediction bias in a Brain-age estimation framework. We applied proposed bias-adjustment scheme coupled by a machine learning-based brain age framework on a large set of metabolic brain features acquired from 675 cognitively unimpaired adults through fluorodeoxyglucose positron emission tomography data as the training set to build a robust Brain-age estimation framework. Then, we tested the reliability of proposed bias-adjustment scheme on 75 cognitively unimpaired adults, 561 mild cognitive impairment patients as well as 362 Alzheimer's disease patients as independent test sets. Using the proposed method, we gained a strong R² of 0.81 between the chronological age and brain estimated age, as well as an excellent mean absolute error of 2.66 years on 75 cognitively unimpaired adults as an independent set; whereas an R² of 0.24 and a mean absolute error of 4.71 years was achieved without bias-adjustment. The simulation results demonstrated that the proposed bias-adjustment scheme has a strong capability to diminish prediction error in brain age estimation frameworks for clinical settings.

Keywords: Bias-adjustment; Brain age; Brain metabolism; Estimation; Pet.

Fig. 1

Example of the age-related bias in brain age delta in our training set, where Δ is the estimated brain age minus the real chronological age. The dashed red line shows the reference line (y = 0), while the solid black line states the regression line. The result of the training set was generated through a 10-fold-cross validation strategy. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 2

First row (A, B and C): Scatter plot of estimated brain age and chronological age on training set followed by three different procedures. The dashed red line shows the identity line (y = x), while the dashed black lines state a 95% prediction band on the model prediction. Second row (D, E and F): brain age delta (Δ: estimated brain age minus chronological age) versus chronological age on training set followed by different procedures. The dashed red line shows the reference line (y = 0), while the dashed black lines state a 95% prediction band on the model prediction. The results of the training set were generated through a 10-fold-cross validation strategy. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

First row (A, B and C): Scatter plot of estimated brain age and chronological age on the independent cognitively unimpaired test set followed by three different procedures. The dashed red line shows the identity line (y = x), while the dashed black lines state a 95% prediction band on the model prediction. Second row (D, E and F): delta age versus chronological age for the independent cognitively unimpaired set after different procedures. The dashed red line shows the reference line (y = 0), while the dashed black lines state a 95% prediction band on the model prediction. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 4

First row (A, B and C): Scatter plot of brain estimated age and chronological age on MCI (orange spot, solid orange regression line) and AD (dark blue spot, solid dark blue regression line) sets followed by different procedures; The dashed red line shows the identity line (y = x). Second row (D, E and F): delta age versus chronological age on MCI (orange spot, solid orange regression line) and AD (dark blue spot, solid dark blue regression line) followed by different procedures; the dashed red line shows the reference line (y = 0). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)