Antonio Maria Chiarelli1, Carlo Sestieri1, Daniele Mascali1, Richard Geoffrey Wise1, and Massimo Caulo1
1Department of Neuroscience, Imaging and Clinical Sciences, University G. D'Annunzio of Chieti Pescara, Chieti Scalo, Italy
Synopsis
We used Machine Learning (ML) to infer gestational age (GA)
at birth, and hence, as a metric of prematurity extent, assess its effect, in
88 premature infants using T2*-w BOLD resting-state connectivity and activity,
and T1-w volume in 90 brain regions. ML was able to infer GA at birth. Analysis of the
spatial distribution of effects indicated that volumetric alterations, in
common with BOLD activity, are partially localized to subcortical structures, but
are associated with widespread alterations of connectivity. Our results suggest
a potential role for ML in early prediction of neurodevelopmental outcome based
on BOLD and anatomical MRI metrics.
Introduction
Recent improvements in neonatal healthcare have reduced the
incidence of brain damage and increased the survival rate of premature
newborns. However, even without evident alteration at standard neuroimaging, this
population is still at risk of poor neurodevelopment1. Identifying improved markers of
altered brain development may help to guide treatment2. Advanced MRI, for example based on functional
BOLD signal, has been used to investigate regional alterations in preterm
brains3. However, the literature reports
discrepancies, e.g. subcortical4 vs. cortical5,6 involvement, plausibly reflecting distinct
a-priori emphasis on brain regions. Machine Learning (ML)7 may overcome analysis biases and it can
account for inter-regional dependencies. We used a ML framework to infer
gestational age (GA) at birth, and assuming it as a metric of prematurity, assess
its effect in different ROIs covering the whole brain, based on T2*-w BOLD resting-state
functional connectivity (rsFC), resting-state functional connectivity density
(rsFCD), functional activity (fractional ALFF, fALFF)8, and T1-w tissue volume.Methods
A total of 88 infants without neurological abnormalities,
born between 25 and 40 weeks of gestational age (GA), (43/88 female, 15/88 >37
weeks) were recruited from the Neonatology Unit of the University Hospital of
Chieti. MRI was performed at term-equivalent age with a 3 T whole-body system
(Achieva3.0 T X-Series) from Philips Healthcare using an 8-channel receiver
coil. Neonates were sedated with 0.05 mg/kg of oral Midazolam. A T1-w sagittal image
(Flip Angle: 8°; TR: 9 ms; TE: 4.2 ms; voxel size: 1×1×1 mm3; FOV: 200×200×150
mm3) and 162, whole-brain, BOLD T2*-w EPI axial volumes (Flip Angle:
90°; TR: 1555 ms; TE: 30 ms; voxel size: 2.5×2.5×3mm3; FOV: 180×180×75 mm3;
slice gap: 0 mm) were acquired at rest. MRI processing, performed using standard FSL9, AFNI10 and ANTs11 tools, is reported in Figure 1. Ninety
subcortical and cortical ROIs were evaluated based on the University of North
Carolina (UNC) infant atlas12 warped to each subject’s brain
anatomy. RsFC matrices were built through pairwise correlations of BOLD signals
among ROIs (4005 independent connections) accounting for whole-brain signal
regression. RsFC density (rsFCD) was evaluated as the square root of the row average of the squared RsFC matrix. fALFF,
which we describe here as ‘BOLD activity’ was evaluated as the ratio between
the BOLD power within 0.01- 0.1 Hz and its total power8 . Regional volume was inferred using deformation
based morphometry (DBM)13. A ML framework was implemented to infer
GA at birth from the different feature spaces (Figure 2). To account for the
large numerosity and collinearity of independent features, a partial least
square (PLS) regression14 was used. In order to optimize the number
of PLS components and to evaluate the algorithm performance, a leave-one-out
nested cross validation (nCV)15 was employed.Results
Figure 3 shows the results of the nCV ML framework used in a
regression (with GA expressed in weeks) and a classification (GA at Birth >
32weeks vs. ≤ 32 weeks) analysis. ML could infer GA at birth relying on the MRI
metrics. Figure 4a-b illustrates the results of performance comparison. RsFC
had better regression performance compared to rsFCD (z=2.0265, p=0.0215) and
the performance of regional volume was higher than that of fALFF (z=2.00,
p=0.0224) and rsFCD (z=3.517, p=2.28·10-4). The comparison of
classification performance indicated a significantly lower performance of rsFCD
compared to both rsFC (z=-2.278, p=0.0114) and regional volume (z=-2.5474,
p=5.5·10-3). Figure 4c shows the difference in performance when using subgroups of
connections or ROIs. Homotopic connections (N=45) appeared to perform less well
than non-homotopic connections (N=3960, z=-2.32, p=0.01), but the effect did
not reach statistical significance when controlling for subgroup numerosity. Figure
3d shows the results obtained for regional metrics divided based on subcortical
(N=8) and cortical (N=82) ROIs. A higher performance was obtained with volume for
cortical with respect to subcortical ROIs (z=2.11, p=0.02). The effect reversed
when controlling for numerosity (z=-1.69, p=0.05). A similar, albeit
statistically weaker, pattern was also observed for fALFF. Discussion
ML was able to infer GA at birth (used as a metric of degree of prematurity) using T2*-w BOLD metrics of brain
activity and connectivity and T1-w metrics of brain volume. ML indicated that
rsFC was the BOLD feature most sensitive to prematurity. The prediction
capability of rsFC vanished when collapsing the information of rsFC into rsFCD.
The results suggested a highly diffused effect of prematurity on rsFC. In contrast, ML indicated that a significant alteration with prematurity was found
for regional volume, albeit not exclusively, in subcortical regions and a
similar pattern was obtained for fALFF. Conclusion
The present study, relying on ML, demonstrated that prematurity is
associated with alterations of T2*-w BOLD functional connectivity and T1-w regional
brain volume, and, to a lesser extent, with modification of BOLD brain
activity. The analysis of the spatial distribution of the effects indicated
that MRI structural alterations, which have some degree of locality involving
subcortical structures, are associated with a widespread effect on BOLD functional
connectivity. The high sensitivity of ML in identifying the complex effects of
GA at birth on multiple MRI metrics of the brain suggests that ML may be a
suitable method for the early prediction of neurodevelopmental outcome based on
BOLD and anatomical MRI metrics.Acknowledgements
This work was partially conducted under
the framework of the Departments of Excellence 2018–2022 initiative of the
Italian Ministry of Education, University and Research for the Department of
Neuroscience, Imaging and Clinical Sciences (DNISC) of the University of
Chieti-Pescara, Italy.References
1. Mento, G. &
Nosarti, C. The case of late preterm birth: sliding forwards the critical
window for cognitive outcome risk. Transl. Pediatr. 4, 214–218
(2015).
2. Rogers,
C. Addressing the Psychosocial Risk Factors Affecting the Developing Brain of
the High-Risk Infant Bringing Science to Law and Policy. Wash. Univ. J. Law
Policy 57, 117–132 (2018).
3. Zhang,
H., Shen, D. & Lin, W. Resting-state functional MRI studies on infant
brains: A decade of gap-filling efforts. NeuroImage 185, 664–684
(2019).
4. Ball,
G. et al. Thalamocortical Connectivity Predicts Cognition in Children
Born Preterm. Cereb. Cortex 25, 4310–4318 (2015).
5. Smyser,
C. D. et al. Prediction of brain maturity in infants using
machine-learning algorithms. NeuroImage 136, 1–9 (2016).
6. Smyser,
C. D. et al. Resting-State Network Complexity and Magnitude Are Reduced
in Prematurely Born Infants. Cereb. Cortex 26, 322–333 (2016).
7. Hastie,
T., Tibshirani, R. & Friedman, J. The Elements of Statistical Learning:
Data Mining, Inference, and Prediction, Second Edition. (Springer-Verlag,
2009).
8. Zou,
Q.-H. et al. An improved approach to detection of amplitude of
low-frequency fluctuation (ALFF) for resting-state fMRI: Fractional ALFF. J.
Neurosci. Methods 172, 137–141 (2008).
9. Jenkinson,
M., Beckmann, C. F., Behrens, T. E. J., Woolrich, M. W. & Smith, S. M. FSL.
NeuroImage 62, 782–790 (2012).
10. Cox,
R. W. AFNI: software for analysis and visualization of functional magnetic
resonance neuroimages. Comput. Biomed. Res. Int. J. 29, 162–173 (1996).
11. Avants, B. B. et al.
A reproducible evaluation of ANTs similarity
metric performance in brain image registration. NeuroImage 54,
2033–2044 (2011).
12. Shi,
F. et al. Infant Brain Atlases from Neonates to 1- and 2-Year-Olds. PLOS
ONE 6, e18746 (2011).
13. Ashburner,
J. et al. Identifying global anatomical differences: Deformation-based
morphometry. Hum. Brain Mapp. 6, 348–357 (1998).
14. Abdi,
H. & Williams, L. J. Partial Least Squares Methods: Partial Least Squares
Correlation and Partial Least Square Regression. in Computational
Toxicology: Volume II (eds. Reisfeld, B. & Mayeno, A. N.) 549–579
(Humana Press, 2013). doi:10.1007/978-1-62703-059-5_23.
15. Filzmoser,
P., Liebmann, B. & Varmuza, K. Repeated double cross validation. J.
Chemom. 23, 160–171 (2009).