Functional connectivity differences of the olfactory network in Parkinson’s Disease, mild cognitive impairment and cognitively normal individuals: A resting-state fMRI study

Abstract

Olfactory dysfunction is an early sign of such neurodegenerative diseases as Parkinson’s (PD) and Alzheimer’s (AD), and is often present in Mild Cognitive Impairment (MCI), a precursor of AD. Understanding neuro-temporal relationships, i.e., functional connectivity, between olfactory eloquent structures in such disorders, could shed light on their basic pathophysiology. To this end, we employed region-based analyses using resting-state functional magnetic resonance imaging (rs-fMRI) obtained from cognitively normal (CN), MCI, and PD patients with cognitive impairment (PD-CogImp). Using machine learning (linear and ensemble learning), we determined whether the identified functional patterns could classify abnormal function from normal function. Olfaction, as measured by objective testing, was found to be most strongly associated with diagnostic status, emphasizing the fundamental association of this primary sensory system with these conditions. Consistently lower functional connectivity was observed in the PD-CogImp cohort compared to the CN cohort among all identified brain regions. Differences were also found between PD-CogImp and MCI at the level of the orbitofrontal and cingulate cortices. MCI and CN subjects had different functional connectivity between the posterior orbitofrontal cortex and thalamus. Regardless of study group, males showed significantly higher connectivity than females in connections involving the orbitofrontal cortex. The logistic regression model trained using the top discriminatory features revealed that caudate was the most involved olfaction-related brain structure (accuracy = 0.88, Area under the Receiver operator characteristic curve of 0.90). In aggregate, our study demonstrates that resting functional connectivity among olfactory eloquent structures has potential value in better understanding the pathophysiology of several neurodegenerative diseases.

Introduction

Among the sensory systems, olfaction is considered the most phylogenetically ancient, having direct anatomical connections with affective and memory systems such as those involving the amygdala and hippocampus (Aqrabawi & Kim, 2020; Pellegrino et al., 2016; Vasavada et al., 2017; Wang et al., 2016). Along with other sensory systems, olfaction is impaired by aging. In fact, hyposmia occurs in over half of individuals between the ages of 65 and 80 and between 62–80 % of those over the age of 80 (Doty et al., 1984). This age-related decline is associated with decreased numbers of cells within the olfactory epithelium (Paik et al., 1992), a decline of olfactory neurons (Rawson et al., 2012), and atrophy within central olfactory system pathways (Dintica et al., 2019; Larsson et al., 2000; Martinez et al., 2017; Murphy, 2002; Su et al., 2015; Wang et al., 2005, 2016).

In addition to aging, per se, olfactory decline is an early sign of such age-related neurodegenerative diseases as Parkinson’s (PD) and Alzheimer’s (AD), and is often present in Mild Cognitive Impairment (MCI), (Cowart, 1989; Doty et al., 1984) showing a positive correlation with cognitive performance (Doty, 2012; Doty & Kamath, 2014; Windon et al., 2020) and a negative correlation with anxiety (Cieri et al., 2023; Krusemark et al., 2013).

In the case of PD, loss of olfactory function is a prodromal biomarker of the disorder, preceding the motor symptoms by more than 4 years (Barresi et al., 2012; Berendse & Ponsen, 2006; Georgiopoulos et al., 2018; Haehner et al., 2007; Tissingh et al., 2001). This is unrelated to the use of dopaminergic medications (Doty et al., 1988). According to the pathology studies of Braak and colleagues (2003), the disease-defining markers of such pathology, e.g., Lewy bodies, first appear within the dorsal medulla and olfactory bulb long before damage is identified within the basal ganglia (Hawkes et al., 1997; Tissingh et al., 2001). Such pathology eventually spreads into limbic and paralimbic regions, including the amygdala, hippocampal gyrus, anterior olfactory nucleus, entorhinal cortex, and piriform cortex (Su et al., 2015).

Using magnetic resonance imaging (MRI), both structural and functional changes have been associated with olfactory deficits in PD. Structural MRI provides basic information about the changes in brain structure and regions, such as alterations in size or shape. Functional MRI (fMRI) detects oxygen level-dependent (BOLD) changes in the MRI signal, capturing time-varying and localized change in neural activity during both resting and stimulus-activated periods. Structural MRI studies have observed some degree of atrophy in PD-related brain regions such as the piriform cortex, amygdala (Wattendorf et al., 2009), parahippocampal gyrus, orbitofrontal cortex (Wu et al., 2011), olfactory bulb, and entorhinal cortex (S. Chen et al., 2014; Su et al., 2015). Findings from fMRI studies have been conflicting. Thus, both hyper- and hypo-activation of olfactory, mediotemporal, and limbic regions have been reported (Morrot et al., 2013; Pellegrino et al., 2016; Vasavada et al., 2017; Wang et al., 2016). Some have noted pathology within the olfactory cortex and its principal projections to limbic, paralimbic, and temporal regions (Moessnang et al., 2011; Su et al., 2015; Welge-Lüssen et al., 2009; Westermann et al., 2008).

In light of the discrepant fMRI findings, there is a need to better understand functional connections (FCs) among olfactory eloquent structures in persons with neurodegenerative disorders that impact cognition. There is also a need to determine how sex influences these connections. Females typically outperform males on olfactory tasks during normal aging, but it is not clear whether this is due to differences in peripheral sensory function or central cognitive processing of olfactory information (Kern et al., 2014, p. 2). One of the possible reasons is that women have higher number of neurons in the olfactory bulb (Oliveira-Pinto et al., 2014). Although this apparent sex advantage, women have also shown a faster age-related reduction in the olfactory cortex volume (Alotaibi et al., 2023). These results suggest that age-related functional changes in males and females may follow a different path and time course, as already suggested (Cieri et al., 2023).

Resting state fMRI (rs-fMRI) provides a means of examining such connections. It is based on the fact that, at rest, the human brain shows synchronous activations between spatially distinct regions, giving rise to resting state networks (RSNs) that are functionally connected.

Interestingly, when olfactory loss is present from birth (congenital anosmia), it does not change functional connectivity (FC) either within or from piriform cortex during rest (Peter et al., 2021). On the other hand, when the same sensory impairment is acquired later in life, it has been associated with a decreased FC within and between the olfactory regions (Iravani et al., 2021; Kollndorfer et al., 2015; Reichert et al., 2018).

In this context, different machine learning techniques have been utilized for a variety of applications in PD (Lötsch & Hummel, 2019). These include preclinical diagnosis of PD using multimodal features including olfaction, cerebrospinal fluid (CSF), and imaging markers (Prashanth et al., 2016); identifying a specific pattern within the University of Pennsylvania Smell Identification Test (UPSIT) score (Doty & Kamath, 2014) that is significantly predictive of PD; and using odor identification as a supportive machine learning based diagnostic tool for PD (Casjens et al., 2013). However, all these studies primarily focused on using the olfaction performance for diagnosis of PD and none of them explored which brain regions were significantly associated to worsening of the odor identification.

Thus, exploring the FC differences of olfactory regions across a wide spectrum of diseases, namely CN individuals, subjects with MCI and PD-CogImp patients, and quantifying the presence of sex differences in these groups, represents a crucial endeavor to unravel the linkage between FC changes in olfaction and the neurodegenerative processes affecting the aging brain.

This study had three main goals. First, to use rs-fMRI to identify the functional connections associated with olfaction in CN individuals, MCI, and PD Cog-Imp. This will be achieved by employing two different models, namely, multiple linear regression (assumes a linear relationship between variables) and Random Forest regression (captures complex nonlinear relationships without assuming linearity). Second, to determine whether the sex of the subjects altered the functional connections. Finally, to determine if the features selected during the feature selection stage of the study can predict an abnormal olfaction score in a binary classification setting utilizing two different machine learning classifiers, one involving a linear decision boundary (logistic regression) and the other a complex nonlinear decision boundary (Random Forest classifier).

Experimental procedures

Material and methods

Participants

This study was authorized by the Cleveland Clinic Institutional Review Board, Study #15–888, and followed the principles outlined in the Declaration of Helsinki. A total of 104 subjects were recruited from the Center for Neurodegeneration and Translational Neuroscience (CNTN). Basic demographical information of the subjects is presented in Table 1. Our inclusion criteria for CN, MCI and PD-CogImp were as follows:

Table 1.

A total of 104 subjects spanning these 3 groups were included in the study. CN represents participants that are cognitively normal, MCI represents the mild cognitively impaired stage. PD-CogImp represents the patients with Parkinson’s disease who also have mild cognitive impairment. B-SIT^® score represents the subjective olfactory evaluation performed using the Brief Smell Identification Test (B-SIT^®) ranging from 0 to 12.

Diagnosis	N	Sex (M/F)	Age	Education	BSIT score
CN	29	11/18	70.48 ± 6.28	16.28 ± 2.56	9.83 ± 1.69
MCI	43	28/15	73.88 ± 5.63	15.70 ± 2.65	8.74 ± 2.17
PD-CogImp	32	18/14	71.09 ± 8.38	14.59 ± 2.95	5.69 ± 3.20

Cognitive normal (CN).

(1) Age between 55 and 90; (2) visual and auditory acuity investigated by physical exam; (3) good general health; (4) score of 26 or greater on the Montreal Cognitive Assessment Test (MoCA; to ensure adequate ability to complete neuropsychological testing); (5) no signs of MCI, PD, or other dementias.

Mild cognitive impairment (MCI).

(1) Subjective memory complaints reported by themselves, study partner, or clinician; (2) objective memory loss defined as scoring below an education-adjusted cut-off score on delayed recall of Story A of the WMS-R Logical Memory Test; (3) global CDR score of 0.5; (4) general cognitive and functional performance sufficiently preserved such that a diagnosis of dementia could not be made at the time of screening.

Cognitively impaired PD (PD-CogImp).

Parkinsonism is defined as bradykinesia, in combination with either rest tremor, rigidity, or both, considering supportive and absolute exclusion criteria (Postuma et al., 2015). The MDS- Unified Parkinson’s Disease Rating Scale (UPDRS) was used; no autopsies were available for the diagnosis confirmation. Cognitive impairment was confirmed by a global CDR score greater than 0.5. All the subjects were screened for psychiatric and neurological conditions at the Cleveland Clinic Lou Ruvo Center for Brain Health and the neuropsychological and olfactory assessments were completed at the same visit. All participants provided written informed consent.

Olfactory evaluation

All participants underwent olfactory examination measured by the Brief Smell Identification Test^® (B-SIT^®) (Doty et al., 1996). The B-SIT^® is a widely used olfactory test, owing to its availability of sex-adjusted normed data and providing a rapid and accurate indication of smell loss. It consists of twelve microencapsulated “scratch and sniff” odorants, with a score range: 0–12, where higher scores represent greater olfactory function. It was developed from the larger 40-item University of Pennsylvania Smell Identification Test (UPSIT) (Doty et al., 1984). During the test, the participants were asked to smell 12 strips containing different odorants, and selectively identify the correct odor from a list of four multiple choices. A B-SIT^® score less than 8 was deemed abnormal irrespective of age and sex.

MRI data acquisition and data processing

The resting state fMRI scans of all subjects were acquired on a Siemens Skyra 3 T scanner with a 32-channel head coil. It involved a total duration of 10 min and 4 s with 850 timepoints (TR=700 ms, TE=28.4 ms, slice thickness = 2.34 mm, flip angle = 40°, echo spacing = 0.61 ms, EPI factor = 96, 64 axial slices in interleaved multi-slice acquisition with a multi-band acceleration factor of 8). In addition, a T1-weighted structural MRI based on a 3D MPRAGE sequence was obtained with the following parameters (TR=2300 ms, TE=2.98 ms, slice thickness = 1 mm, flip angle = 9°).

The EPI data were not slice time corrected because of the short TR (700 ms) and a multiband factor of 8. The data were then realigned to the mean EPI image using Statistical Parametric Mapping version 12 (https://www.fil.ion.ucl.ac.uk/spm/) followed by distortion correction using topup in FSL (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/topup) to correct for distortions in the phase encoding direction (AP) using data from two different phase encoding directions. These were then coregistered to the subject’s T1 image and then normalized to the MNI-152 2 mm template using Advanced Normalization Tools (ANTs) software (https://stnava.github.io/ANTs/). The temporal preprocessing steps included linear detrending based on Discrete Cosine Transforms to remove signal instabilities less than 0.01 Hz and variance normalization. No explicit motion regression was performed to avoid introducing high frequency artifacts (J. E. Chen et al., 2017).

Functional connectivity and regression analysis

A total of 30 regions of interest (ROIs, 15 for each hemisphere) considered relevant to olfaction were derived from the 120 ROIs of the AAL2 atlas (E. T. Rolls et al., 2015). The 30 ROIs spanned 9 brain regions (Fig. 1). The average time series was extracted from each of these ROIs and Pearson’s correlations based functional connectivity were calculated for all pairwise ROI combinations. This was repeated for each subject across all three groups (CN, MCI, and PD-CogIMP) resulting in a total of 435 unique functional connections organized into 30 × 30 subject specific FC matrices.

Fig. 1.

The cumulative list of the 9 ROIs that were formed from the 30 ROIs derived from the AAL2 atlas. Each of the 15 ROIs under the column (AAL2 derived ROI name) have left and right hemisphere counterparts resulting in a total of 30 ROIs. The corresponding locations of these ROIs overlaid on a standardized MNI template is also shown. All these ROIs were derived from the Automated Anatomical Labelling Atlas (AAL2 atlas).

The functional connections involving a particular ROI (left and right hemispheres combined) alongside age, diagnosis, sex, and education were used as the predictors and were subsequently modeled to understand how they influence the B-SIT^® score (target variable), representing olfaction. Two different approaches based on the underlying relationship between the predictors (functional connections and the demographics variables) and the B-SIT^® score were implemented. The first one is a multiple linear regression model that operates by minimizing the residual sum of squares to estimate the model parameters (the coefficients of the predictors). Unlike the former that assumes linearity in the parameters, the second model, Random Forest regression is devoid of any prior assumptions and outputs an average prediction from an ensemble of de-correlated trees.

A stepwise multiple linear regression model based on bidirectional elimination was used to shortlist the statistically significant predictors that influence the B-SIT^® score based on a linear relationship assumption. This was achieved by determining the associated p-value of the F statistic of the model, both with and without each of the predictors and including only those predictors that reject the null hypothesis (predictor having zero coefficient) at a 95 % significance level in the final model.

Conversely, the Random Forest regression model is an ensemble method which was hyperparameter tuned for the number of trees (between 500 and 1000), learning rate (between 0.001 and 0.1), and minimum leaf size (between 1 and 50) using 30 runs of Bayesian optimization to identify the features that were related to the B-SIT^® score without assuming any underlying relationship trends. A recursive feature elimination scheme starting with a full set and iteratively eliminating the least important features was implemented. This yielded an optimal set of features that had the least mean squared error across all the iterations. The feature importance, termed permutation feature importance, was calculated by randomly permuting each of the predictors in the out-of-bag samples and evaluating the difference in the model error due to this permutation.

Both these regression models were repeated for functional connections involving each ROI (left and right hemispheres combined) separately thus evaluating the influence of functional connectivity between different brain regions and the subject demographics on olfaction.

Classification analysis

A binary classifier model was then built to predict an abnormal B-SIT^® score using the top features obtained from the regression analysis. All scores ≤ 8 were considered abnormal and coded 1 and the rest coded as 0. To evaluate the predictive capability of each of the ROIs, the 30 ROIs were consolidated into 9 ROIs by considering ROIs belonging to different hemispheres or segments of the same brain region as a single ROI (see Table 2 for more information).

Table 2.

The results of the parametric two-way ANOVA for the continuous demographics variables (Age, education and the B-SIT^® score) showing the F statistic and the associated p-values. The significant comparisons (p-value < 0.05) at 95 % significance level are highlighted in bold. A critical F-statistic value of 3.08 based on df1 = 2 and df2 = 101 at α = 0.05 was chosen as the threshold where df1 represents degrees of freedom in the numerator (number of groups – 1) and df2 represents the degrees of freedom in the denominator (Total number of samples – number of groups).

	Variable	Factor 1 (Diagnosis)		Factor 2 (Gender)
		Test statistic	p-value	Test statistic	p-value
Two-way ANOVA	Age	2.04	0.14	1.74	0.19
	Education	3.68	0.03	4.31	0.04
	BSIT score	23.90	3.52E-09	0.27	0.60

Two different classifier models, namely a logistic regression model featuring a linear decision boundary was built using the features selected by multiple linear regression whereas an ensemble algorithmic based Random Forest classifier was built using the features selected by the corresponding Random Forest regression model. The logistic regression model included a L2 regularization to prevent overfitting by including an additional squared magnitude of the coefficients in the loss function. The performance of the classifier in predicting an abnormal B-SIT score were assessed using multiple criteria; (a) accuracy (proportion of the correct predictions to the total number of predictions), (b) precision (ratio of true positives to the total number of true and false positives), (c) recall (ratio of the true positives to the total number of true positives and false negatives) and (d) The Area under the Receiver Operating Characteristic (ROC) curve (AUC-ROC) quantifies the overall ability of a binary classification model to distinguish between positive and negative classes, with higher values indicating better model performance.

Statistical analysis

Parametric statistical test based on a two factor Analysis of Variance (ANOVA) was used to compare the means of the continuous demographic variables across sex and the diagnostic status. Since we also wanted to evaluate if the functional connectivity features identified during the feature selection stage showed any significant differences across both factors (sex and the diagnostic status), we implemented a similar two-way ANOVA to evaluate the independent effects of these factors on the connectivity. We also included an interaction term (interaction between sex and diagnosis) to identify if sex modulates the association between connectivity and the disease state. The uncorrected p-values obtained were then corrected for multiple comparisons (number of top features within each approach × 3) using the Benjamini-Hochberg procedure. This was followed by the post-hoc Tukey multiple comparisons test (Keselman & Rogan, 1977) to identify the significant differences across all pairwise combinations for the diagnostic factor. Since the other factor (sex) had only 2 levels, no further post-hoc analysis was done.

All the steps involved in functional connectivity analysis, linear regression and the random forest methods were implemented using in-house scripts programmed in MATLAB Release 2021b, The Math-Works, Inc., Natick, Massachusetts, United States (www.mathworks.com). The logistic regression was carried out using scikit-learn (Buitinck et al., 2013) using Python 3.11 and the statistical analysis was performed using the SciPy library (Virtanen et al., 2020).

Results

Demographics

The results of two-way ANOVA with 2 factors (diagnosis, sex) are included in Table 2. With respect to the continuous demographic variables, we found education to be significantly different across both sex (p = 0.04) and diagnosis (p = 0.03) whereas the B-SIT^® score was found to be significantly different only across diagnosis (p = 3.52e-09). However, age was identified to be not significant across any of the factors.

Functional connectivity and regression

Multiple linear regression

Only 2 functional connections with associations with olfaction were identified to show significant differences based on the 2-way ANOVA, one for each of the 2 factors (sex and diagnosis) as seen in Table 3. Interestingly, both these features involved the orbital frontal cortex (Table 3). With regards to sex, the males showed significantly increased connectivity than females (multiple comparisons corrected p = 9.e-3) for the FC between the left anterior cingulate cortex and the left lateral orbital frontal cortex (Fig. 2B). In terms of the differences based on diagnosis, a significantly lower FC was observed in the PD-CogImp group compared to both CN (p = 0.002) and MCI (p = 0.006) with the former being more pronounced. No significant differences were found between CN and MCI. Moreover, no significant interaction effects between sex and diagnosis for any of the FC features.

Table 3.

The F statistic and p-values based on the parametric two-way Analysis of Variance (ANOVA) with diagnosis and sex as the factors along with their interaction effect are shown. The uncorrected p-values and their adjusted versions after the Benjamini-Hochberg procedure are also shown. The features shown here are the features that had significant uncorrected p-values at a 95 % significance level obtained from the combined list of functional connectivity features shortlisted from stepwise linear regression. A critical F-statistic value of 3.08 based on df1 = 2 and df2 = 101 at α = 0.05 was chosen as the threshold where df1 represents degrees of freedom in the numerator (number of groups – 1) and df2 represents the degrees of freedom in the denominator (Total number of samples – number of groups).

Feature	Main effect	F-statistic	p-value (uncorrected)	p-value (corrected)
Anterior cingulate cortex (L), Lateral orbital frontal cortex (L)	Sex	17.55	6.1e-5	9.1e-3
Medial orbital frontal cortex (L), thalamus (L)	Diagnosis	7.93	6.4e-4	0.047

Fig. 2.

Violin plots showing the distribution of functional connectivity across the two factors, (A) diagnosis and (B) sex used in the two-way ANOVA model with the p-values corrected for multiple comparisons. The left figure (A) shows the unique features that show significant differences across diagnosis along with the p-values from the post-hoc Tukey’s Honestly Significant Difference (HSD) test all pairwise comparisons (CN vs MCI, MCI vs PD-CogImp and CN vs PD-CogImp). The right figure (B) shows the unique features that show significant differences across sex along with the p-value corrected using the Benjamini-Hochberg procedure for the comparison (male vs female). For both figures (A) and (B), the features were derived from the combined list of top features selected by the stepwise linear regression model.

Random Forest regression

The following hyperparameter values were identified (number of trees = 1000, learning rate = 0.01 and minimum leaf size = 15). Similar to the trends identified with the features selected via linear regression, the connections showing significant sex differences involved the orbital frontal cortex (Table 4, Fig. 3B). Also, the connection between left insula and the lateral orbital frontal cortex was significantly affected by both sex (p = 0.037) and diagnosis (p = 0.048) as seen in Table 4. Fig. 3 shows that males showed significantly increased connectivity than females (p < 0.05) whereas no significant interaction effects between diagnosis and sex were found. The PD-CogImp group had the lowest FC among all three groups (Fig. 3A) with significant differences observed between CN and PD-CogImp for all the features (p < 0.05 for CN vs PD-CogImp in Fig. 3A). The connections involving the orbital frontal cortex showed a significantly lower FC in the PD-CogImp compared to MCI (p ≤ 0.05) whereas the connection between insula and the ACC showed no significant differences. Irrespective of the regression methodology applied, diagnosis variable was found to have the most significant influence on the B-SIT score for all the ROIs chosen suggesting that disease state (CN or MCI or PD-CogImp) had the most influence on the B-SIT score in both the regression approaches (highest value of T-statistic for the stepwise multiple linear regression and the highest permutation feature importance among all predictors) as seen in Table 5.

Table 4.

The F statistic and p-values based on the parametric two-way Analysis of Variance (ANOVA) with diagnosis and sex as the factors along with their interaction effect are shown. The uncorrected p-values and their adjusted versions after the Benjamini-Hochberg procedure are also shown. The features shown here are the features that had significant uncorrected p-values at a 95 % significance level obtained from the combined list of functional connectivity features shortlisted from random forest regression. A critical F-statistic value of 3.08 based on df1 = 2 and df2 = 101 at α = 0.05 was chosen as the threshold where df1 represents degrees of freedom in the numerator (number of groups – 1) and df2 represents the degrees of freedom in the denominator (Total number of samples – number of groups).

Feature	Main effect	F-statistic	p-value (uncorrected)	p-value (corrected)
Anterior cingulate cortex (L), Lateral orbital frontal cortex (L)	Sex	17.55	6.1e-5	9.1e-3
Anterior cingulate cortex (R), Lateral orbital frontal cortex (L)	Sex	11.52	9.98e-4	0.045
Insula (L),Lateral orbital frontal cortex (L)	Sex	10.84	1.39e-3	0.037
Insula (L),Anterior cingulate cortex (R)	Diagnosis	7.22	1.19e-3	0.040
Anterior cingulate cortex (R), Lateral orbital frontal cortex (L)	Diagnosis	6.32	2.62e-3	0.044
Insula (L),Lateral orbital frontal cortex (L)	Diagnosis	5.98	3.54e-3	0.048

Fig. 3.

Violin plots showing the distribution of functional connectivity across the two factors, (A) diagnosis and (B) sex used in the two-way ANOVA model with the p-values corrected for multiple comparisons. The left figure (A) shows the unique features that show significant differences across diagnosis along with the p-values from the post-hoc Tukey’s Honestly Significant Difference (HSD) test all pairwise comparisons (CN vs MCI, MCI vs PD-CogImp and CN vs PD-CogImp). The right figure (B) shows the unique features that show significant differences across sex along with the p-value corrected using the Benjamini-Hochberg procedure for the comparison (male vs female). For both figures (A) and (B), the features were derived from the combined list of top features selected by the random forest regression model.

Table 5.

The T-statistic values and the permutation feature importance values for the most important predictor (highest T-statistic in multiple linear regression and highest permutation feature importance) for each of the cumulative ROI is shown for multiple linear regression and random forest regression, respectively. Consistently diagnosis was observed to be the predictor that was most significantly associated with olfaction score. The negative T-statistic here shows that olfaction score decreases as you go from CN to MCI and PD-CogImp. Permutation importance shown here is the mean difference in model error after permutation of the diagnosis feature and before the permutation.

Brain region	Multiple linear regression		Random forest regression
	Most important predictor
	T-statistic	Predictor name	Permutation importance	Predictor name
Amygdala	−7.24	Diagnosis	1.91	Diagnosis
Caudate	−6.62	Diagnosis	2.41	Diagnosis
Cingulate cortex	−6.83	Diagnosis	1.19	Diagnosis
Frontal orbital (inferior)	−7.02	Diagnosis	2.29	Diagnosis
Hippocampus	−8.24	Diagnosis	1.42	Diagnosis
Insula	−6.17	Diagnosis	2.94	Diagnosis
Orbitofrontal cortex	−8.52	Diagnosis	1.07	Diagnosis
Olfactory cortex	−7.02	Diagnosis	2.24	Diagnosis
Thalamus	7.26	Diagnosis	2.67	Diagnosis

Machine learning classifier

The classifier performance evaluated using accuracy, precision, recall and AUC-ROC shows that the logistic regression model with L2 regularization performs better than the Random Forest classifier model for all the 9 cumulative ROIs. This can be verified from Fig. 4 and Table 6. Cumulatively, the classifier built using the top features involving caudate, thalamus and the inferior frontal orbital cortex from the model with a linear decision boundary showed the maximum accuracies (0.88, 0.84 and 0.84 respectively) and correspondingly maximum AUC-ROC values of 0.90, 0.89 and 0.90 respectively (Table 6). Both the classifier models consistently had slightly higher precision values compared to recall suggesting that the model prioritizes reducing the number of false positives (normal olfaction scores incorrectly classified as abnormal) as seen from Table 6.

Fig. 4.

Binary classification results for predicting abnormal B-SIT^® score (scores ≤ 8) using the top features selected by (A) stepwise multiple linear regression and (B) random forest for the 9 cumulative ROIs are shown. The AUC results are ordered from highest to lowest and the corresponding mean AUC values are highlighted in bold. The left figure (A) shows AUC-ROC curve (false positive rate vs true positive rate) for the testing set samples for the logistic regression model based on L2 regularization whereas the right figure (B) shows the same curve for the out of bag samples in the random forest model.

Table 6.

The different metrics (accuracy, precision, recall and Area under the Receiver Operating Characteristic curve (AUC-ROC)) used in the evaluation of the machine learning classifier have been summarized for both logistic regression and random forest models. The regions and the corresponding metrics are ordered based on the maximum AUC-ROC values observed for the logistic regression model.

Brain region	Logistic regression				Random forest
	Accuracy	Precision	Recall	AUC-ROC	Accuracy	Precision	Recall	AUC-ROC
Caudate	0.88	0.90	0.83	0.90	0.73	0.80	0.75	0.79
Frontal orbital (inferior)	0.84	0.95	0.82	0.90	0.81	0.88	0.80	0.82
Amygdala	0.81	0.89	0.81	0.90	0.75	0.85	0.75	0.81
Orbitofrontal cortex	0.78	0.89	0.77	0.89	0.74	0.85	0.74	0.79
Thalamus	0.84	0.95	0.82	0.89	0.77	0.80	0.80	0.84
Olfactory cortex	0.75	0.79	0.79	0.89	0.76	0.87	0.75	0.83
Cingulate cortex	0.81	0.95	0.78	0.88	0.77	0.88	0.76	0.83
Hippocampus	0.78	0.84	0.80	0.88	0.74	0.85	0.74	0.76
Insula	0.78	0.89	0.77	0.87	0.82	0.88	0.82	0.84

Discussion

A major finding of our study was the discovery of a consistent pattern of lower FC in the PD-CogImp compared to the CN group, especially when regions like insula, OFC, ACC, and thalamus were involved. This pattern was also evident between MCI and PD-CogImp when we considered the OFC connected to the thalamus, insula or the ACC. These results are not surprising since all these regions are involved in the perception of odors, with their emotional and memory associated.

Any FC involving both cingulate cortex and the OFC showed significantly lower value in the PD-CogImp group regardless of the specific sub-region or hemisphere involved. This result confirms the importance of these two regions in the olfaction process. Particularly, the role of ACC in sensory processing, encoding and attentional control has been well studied (Cera et al., 2019; Mayberg, 1997). The ACC also influences the anterior olfactory nucleus and has been proposed to activate the entire primary olfactory cortex to mediate the process of rapid attention to olfactory stimuli (García-Cabezas & Barbas, 2014). This view is consistent with our results of lowered FC in PD Cog-Imp compared to the CN, with the MCI group positioned in the middle of this trajectory. This interpretation gains further credibility when examining the results, which reveal a notably diminished FC between the insula and the thalamus. The insula is a pivotal structure especially involved in multisensory and affective processing, as well as social functions like empathy (Gogolla, 2017). This is one of the reasons why olfaction, emotion and social functions can be closely interrelated since the olfactory stimulus is directly connected to the olfactory cortex without passing through the thalamus. Nonetheless, following the connection to the primary olfactory cortex, the stimulus proceeds to the thalamus, specifically to the mediodorsal thalamic nucleus (Courtiol & Wilson, 2015) which explains the lower FC between the ACC and the thalamus. Similarly, the OFC results are also not surprising, given that unimodal olfactory and unimodal gustatory neurons are located in this region (E. Rolls & Baylis, 1994; E. T. Rolls, 2004).

Another interesting result concerns the difference between males and females. Males showed significantly increased FC between the ACC and the lateral OFC compared to females. The sex of individuals plays an important role in determining olfactory abilities, in terms of sensitivity, identification, familiarity, and recognition of odors (Brand & Millot, 2001). Studies aimed specifically at testing sex differences in olfaction generally obtained results in favor of females (Cain, 1982; Doty et al., 1985; Koelega & Köster, 1974). Although studies involving bigger samples suggested that smell detection ability (Kern et al., 2014) or olfactory identification (Sorokowska et al., 2015) are similar between males and females, a general better olfaction ability of the latter, in comparison to males, are taken for granted. In fact, many reviews do not focus on existence of such a difference, but rather try to determine its cause (Brand & Millot, 2001; Doty & Cameron, 2009; Sorokowska et al., 2015). Since this sex difference was found across our three different diagnostic groups, is difficult to interpret, but it is in general consistent with our previous results, when we used the graph theory metrics approach on a different sample (e.g., degree centrality, global efficiency, local efficiency, etc.) showing lower values for CN (Cieri et al., 2021) and MCI (Yang et al., 2022) in females compared to males.

Finally, from a more methodological perspective, the logistic regression model performed better than the Random Forest classifier in PD-CogImp diagnosis as evident in the former’s consistently higher AUC-ROC values for all the brain regions. This result was potentially due to our relatively small sample size, our low variance within it, lower number of features (top features set) used during classification and the possibly predominant linear relationship of the data resulting in the logistic regression model (with a linear decision boundary) capturing this trend better than the Random Forest model (with a complex nonlinear decision boundary) that is less interpretable and requires a more extensive model tuning (hyperparameters).

In this study we have shown that a distinct olfactory functional network pattern may represent a change involved in PD which deserves more investigation. This research of FC of structures within the olfactory system network suggests the potential utility of such measures in identifying pathophysiological mechanisms. Given that the loss of olfaction is one of the first symptoms of pathological aging, FC investigation of olfactory-related regions can be helpful to better understand the early signs of hyposmia present in PD, anticipating the diagnostic process and guiding the development of new potential therapy.

Our study has few limitations, such as a small sample size and the absence of a longitudinal design. Moreover, the study would have been more comprehensive if a group consisting of PD without cognitive impairment were included which would’ve enabled better characterization of the olfactory dysfunction differences specifically related to the PD and cognitive impairment.

Our study has also notable strengths, particularly given the scarcity of research in this area. We added a significant finding in the realm of olfaction dysfunction and functional connectivity. In fact, a distinct olfactory FC pattern may represent a change involved in PD. FC of the olfactory network highlights the potential of such measures in identifying pathophysiological mechanisms, with potential role in the diagnostic process. We employed a more granular analysis, identifying the features that most significantly influence olfaction, but also underlined that FC features are capable of predicting the B-SIT^® score. We also underscored the importance of sex differences in this investigation and in general in the aging process. Finally, we utilized models both with and without a linearity assumption between the predictors and the B-SIT^® score, multiple linear regression and Random Forest regression respectively.

Olfactory deficits are one of the earliest features of AD and PD. It is important to further investigate whether neural changes in terms of FC could anticipate the hyposmic process. The association of the B-SIT^® with the rs-fMRI could shed new light, focusing on network integrity rather than focal pathology, which may be beneficial in understanding PD pathology. Olfaction and its functional connectivity should also be investigated taking into account potential differences between males and females. More studies are needed to explore discriminatory connectivity signatures of olfaction and its impairment in dementia such as PD, driving clinically important network differences that may influence the monitoring and treatment of PD patients.