The metabolic spatial covariance pattern of definite idiopathic normal pressure hydrocephalus: an FDG PET study with principal components analysis

Idiopathic normal pressure hydrocephalus (iNPH) is clinically characterized by movement and cognitive disturbance as well as urinary incontinence, forming the typical “Hakim triad” [1]. Other neurodegenerative conditions such as Alzheimer’s disease (AD) or Parkinson’s disease (PD) can present similar clinical symptoms and thus hamper differential diagnoses. Since symptoms in iNPH are potentially reversible, e.g., by ventriculoperitoneal cerebrospinal fluid (CSF) shunting, identification of patients with iNPH is essential [2].

Magnetic resonance imaging (MRI)-based signs of iNPH are incorporated in current diagnostic criteria of iNPH [3]. The established pattern on MRI includes enlargement of the ventricles and the Sylvian fissure and constricted cerebrospinal fluid spaces in the high convexities of the brain (“tight high convexity”) leading to the so-called disproportionately enlarged subarachnoid space hydrocephalus (DESH) and a decreased callosal angle [4‐7].

In the workup of suspected iNPH expert opinion recommends FDG PET for the detection of typical patterns of potentially co-existing neurodegenerative diseases which might cause part of the symptoms [8]. Moreover, there is evidence of altered cerebral metabolic activity in patients with iNPH: Previous molecular imaging studies with positron emission tomography (PET) and [¹⁸F]fluorodeoxyglucose (FDG) showed decreases in glucose metabolism in iNPH either wide-spread [9] or accentuated in the vicinity of the lateral ventricles [10]. Additionally, Ohmichi and colleagues [11] used perfusion single-photon emission computed tomography (SPECT) to describe the so-called convexity apparent hyperperfusion (CAPPAH) sign, i.e., increased cerebral blood flow at the high convexity.

Considering previous observations [10, 11], we applied scaled subprofile model principal components analysis (SSM-PCA) to examine whether there is an iNPH-related spatial covariance pattern (iNPHRP) that would comprise regions with both relative increased and decreased weight of cerebral glucose metabolism. SSM-PCA is a powerful voxel-wise analysis method to reduce the dimensionality of FDG PET data that has previously been applied to study various diseases like parkinsonian syndromes [12‐14], dementia [15, 16] as well as non-neurodegenerative disorders [17]. It allows to determine the degree of similarity between a metabolic spatial covariance pattern and each individual’s FDG PET data resulting in a one-score-per-subject rating. We hypothesize that such a pattern and patients’ scores would not only distinguish iNPH from healthy controls, but also from other neurodegenerative conditions like AD or PD, which is of high clinical relevance.

We identified 11 patients with definite iNPH (CSF shunt responders) and 34 patients with possible iNPH (DESH and more than one clinical symptom of the Hakim triad, see Table 1).

The SSM-PCA group analysis identified 59 principal components among which principal component 1 accounted for 33.69% of the variance in the data and was the only component that showed a group difference between patients with definite iNPH and HEC (p < 0.001, AUC ROC = 0.99). The iNPHRP includes voxels with positive voxel weights (corresponding to the relative increase of glucose metabolism, see Online Resource 2) mainly at the paracentral lobule and negative voxel weights (relative metabolic decreases, see Online Resource 2) at the vicinity of the lateral ventricles including the caudate nucleus, thalamus, and the cingulum (Fig. 1).

Pairwise Mann–Whitney U tests showed that the resulting iNPHRP scores in definite iNPH were significantly higher than in AD and PD (p < 0.05, Bonferroni-corrected for multiple comparisons). Scores in definite iNPH were also numerically higher than in possible iNPH, although they failed to reach significance after Bonferroni correction. In possible iNPH, iNPHRP scores were significantly higher than in AD and PD (Bonferroni-corrected p < 0.05; see Fig. 2, Table 1, and Online Resource 1).

When the ROC-derived cut-off (iNPHRP score ≥ 26.57) was used as a threshold, the iNPHRP score discriminated definite iNPH from AD and PD with an accuracy of 96% and 100%, respectively. Using the same threshold, iNPHRP scores also discriminated possible iNPH from AD (accuracy of 83%) and PD (accuracy of 86%; please see Online Resource 1).

The confounder analysis with covariates “age,” “sex,” and “PET/CT system” essentially replicated these results: the main effect of factor “group” was highly significant (F(4,158) = 87.0, p < 0.001), with significantly higher iNPHRP scores in definite iNPH than in possible iNPH (adjusted p < 0.005), AD, PD, and HEC (adjusted p < 0.001) and significantly higher iNPHRP scores in possible iNPH than in AD, PD, and HEC (all adjusted p < 0.001). Partial eta² revealed that the group factor explained more variance in the data (partial eta² = 0.69) than the other variables (age, partial eta² = 0.08; sex, partial eta² = 0.02; PET/CT system, partial eta² = 0.03).

Exploratory post hoc analyses of the relationship between iNPHRP scores and clinical features of iNPH showed that iNPH patients with cognitive impairment had higher iNPHRP scores (z = 2.77 + / − 1.46) than those without (2.03 + / − 0.04; p = 0.002, t test). In patients with definite or possible iNPH, higher iNPHRP scores were associated (at trend level) with more time needed to accomplish the timed-up-and-go test (r = 0.45, p = 0.082, available in 16/45 iNPH patients), but not 15 m walking (N = 18/45) or the five-chair test (N = 13/45).

We determined the metabolic spatial covariance pattern of definite iNPH on FDG PET in comparison with healthy elderly controls. Subject scores of iNPHRP were significantly higher in definite iNPH than in AD and PD patients, allowing for a discrimination between the groups with an accuracy of 96% and 100%, respectively, which is of high differential diagnostic importance.

The constructed metabolic covariance pattern of definite iNPH comprises regions with relative increases as well as decreases of metabolism (see Online Resource 2). We assume that the relative hypometabolism around the lateral ventricles mainly reflects the typical ventriculomegaly in iNPH and therefore not an altered glucose metabolism per se, but rather an altered ratio of brain parenchyma to CSF. Hence, less regional FDG uptake at the caudate nucleus might rather be due to this partial volume effect. Therefore, we assume that other imaging modalities, like DAT SPECT, might also detect abnormalities in this region [26]. We interpret the relative hypermetabolism in the paracentral midline region in iNPH to be caused by a mixture of partial volume effects, i.e., the narrowing of the sulci and subarachnoid space over the high-convexity/midline surface on one hand, and on the other hand relatively preserved neuronal function in this area in the presence of impaired neuronal function in other areas of the brain. Hence, the relative hypermetabolism in this region in our view does not indicate increased glucose metabolism of brain parenchyma. Taken together, the metabolic covariance pattern of definite iNPH parallels the morphologic changes in definite iNPH with an emphasis on the ventricular enlargement and high-convexity/midline tightness (as corroborated by MRI data in Online Resource 3). This pattern might, beyond the morphologic changes, also reflect true metabolic changes, e.g. in the caudate nucleus. Caudate hypometabolism has previously been reported in iNPH [10], and the caudate nucleus has been linked to typical iNPH symptoms such as gait disturbance, cognitive impairment, and apathy [3].

Decreased cerebral glucose metabolism in iNPH has been described before: globally [27, 28] and in the vicinity of the lateral ventricles, comprising the striatum [10, 29]. By contrast, increased metabolism in iNPH was only reported in a volume-of-interest (VOI) based analysis [29]: The authors observed higher FDG uptake compared to healthy controls in a medial parietal VOI, but only in asymptomatic individuals with signs of iNPH on CT, but not in the group of patients with imaging and clinical features of iNPH. We postulate that increased FDG uptake could be observed in the medial parietal VOI due to its close proximity to the paracentral lobule, which was identified as the region of maximum positive weight (relatively increased metabolism) on the iNPHRP in our study. As VOI-based approaches rely on a priori assumptions, information might be lost by omitting certain regions in contrast to voxel-based approaches. Results of our whole-brain voxel-wise analysis strongly suggest to regard relative hypermetabolism of the paracentral region as an FDG PET indicator of iNPH.

Relative metabolic increase in the paracentral region closely resembles the CAPPAH sign on perfusion SPECT [11], though the current study introduces substantial improvement and novelty: (A) It provides a whole-brain, voxel-wise pattern, which can be expressed in a single numeric value per patient. (B) The iNPHRP also comprises relative decreases at the lateral ventricles, whereas the CAPPAH sign consists only of the hyperperfusion at the high convexity. (C) The iNPHRP was statistically derived in an observer-independent, data-driven fashion from definite iNPH patients. By contrast, the CAPPAH sign was presented after its observation in clinical routine, and the higher prevalence of iNPH in individuals with the CAPPAH sign was only secondarily confirmed. (D) The iNPHRP achieved a slightly better discrimination of iNPH patients from healthy elderly controls (sensitivity of 100% (11/11) and specificity of 96% (46/48)) compared to the CAPPAH’s sensitivity of 80% (24/30) and specificity of 100% (19/19)). (E) The SSM-PCA-based approach permits the calculation of individual subject scores of the iNPHRP as continuous values whereas the visual read of the CAPPAH sign only allows dichotomous classification. Thus, iNPHRP scores may be particularly suitable for prediction models and follow-up examinations. The iNPHRP score might hence convey information about disease severity, which needs to be addressed by further studies. (F) Finally, implementation of the SSM-PCA methodology and read-out of disease-related metabolic covariance pattern can easily be implemented in clinical routine and does not depend on rater experience, and therefore may assist less experienced raters. Of note, for clinical purposes, it may be advisable to read out and combine the spatial covariance patterns of several diseases of interest (e.g., iNPH-, PD-, and AD-related patterns), which has been shown to be a very powerful diagnostic approach in other clinical scenarios such as differential diagnosis of parkinsonian syndromes [14].

Individual iNPHRP scores could be calculated not only in patients from the derivation sample (definite iNPH), but, importantly, also in patients from other samples with overlapping symptoms (e.g. AD, PD). With diagnostic accuracies for iNPH of 100% against PD and 96% against AD, we strongly advocate for the use of FDG PET in the workup of neurodegenerative conditions to assess the presence of the metabolic iNPH spatial covariance pattern. This underlines its potential for clinical application in individual cases.

All iNPH patients analyzed in this study had DESH. Therefore, the identified iNPHRP might apply to iNPH with DESH specifically. According to current guidelines, a substantial proportion of definite iNPH might not have DESH [3]. Whether the iNPHRP is sufficiently sensitive to detect non-DESH iNPH needs to be addressed by further studies. Still, it might be considered useful in clinical routine to have the iNPHRP as an additional DESH marker, as DESH is among the best presurgical neuroimaging predictors of shunt response and to date only assessed dichotomously and subjectively [3, 30]. The possible iNPH group showed overlapping, but still numerically lower average iNPHRP scores than definite iNPH. Due to the retrospective character of this study, the further course of the patients in the possible iNPH group could not be systematically assessed. Therefore, a potential association between iNPHRP scores and shunt responsiveness in these could not be evaluated. This needs to be addressed in further studies, together with the reversibility of the observed spatial covariance pattern and its correlation with clinical symptoms’ expression. Other limitations of this study include the small sample size of definite iNPH cases and a lack of control groups with mixed diseases (e.g., AD plus iNPH). Moreover, the AD and PD groups were younger than the iNPH group, though age did not contribute substantially to the group differences in iNPHRP scores. A final limitation of this retrospective study is the sparse clinical data characterizing the iNPH patients: While documented data were missing in many cases, we did observe a relationship between higher iNPHRP scores and cognitive as well as motor impairment.

We defined a novel metabolic spatial covariance pattern of iNPH that might facilitate the differential diagnosis of iNPH versus other neurodegenerative disorders. Further validation in prospective cohorts is needed, including correlation studies with clinical features, possible prediction of shunt response, and pattern reversibility.