Brain Age Is Associated with Cognitive Outcomes of Cholinesterase Inhibitor Treatment in Patients with Mild Cognitive Impairment

Abstract

Background:

The effect of cholinesterase inhibitor (ChEI) on mild cognitive impairment (MCI) is controversial. Brain age has been shown to predict Alzheimer’s disease conversion from MCI.

Objective:

The study aimed to show that brain age is related to cognitive outcomes of ChEI treatment in MCI.

Methods:

Brain MRI, the Clinical Dementia Rating (CDR) and Mini-Mental State Exam (MMSE) scores were retrospectively retrieved from a ChEI treatment database. Patients who presented baseline CDR of 0.5 and received ChEI treatment for at least 2 years were selected. Patients with stationary or improved cognition as verified by the CDR and MMSE were categorized to the ChEI-responsive group, and those with worsened cognition were assigned to the ChEI-unresponsive group. A gray matter brain age model was built with a machine learning algorithm by training T1-weighted MRI data of 362 healthy participants. The model was applied to each patient to compute predicted age difference (PAD), i.e. the difference between brain age and chronological age. The PADs were compared between the two groups.

Results:

58 patients were found to fit the ChEI-responsive criteria in the patient data, and 58 matched patients that fit the ChEI-unresponsive criteria were compared. ChEI-unresponsive patients showed significantly larger PAD than ChEI-responsive patients (8.44±8.78 years versus 3.87±9.02 years, p = 0.0067).

Conclusions:

Gray matter brain age is associated with cognitive outcomes after 2 years of ChEI treatment in patients with the CDR of 0.5. It might facilitate the clinical trials of novel therapeutics for MCI.

Keywords

Alzheimer’s disease brain age cholinesterase inhibitor clinical dementia rating cognitive outcome magnetic resonance imaging mild cognitive impairment mini-mental state examination

INTRODUCTION

As the world marches toward an aging society, the population of patients with dementia due to aging is increasing. As of 2019, there were approximately 57.4 million people living with neurodegenerative dementia, and the population was estimated to reach 152.8 million in 2050 [1]. The increasing population of dementia poses tremendous economic burdens to patients’ families and healthcare systems worldwide. Pharmacological or non-pharmacological treatments have been actively developed to modify the disease or delay the progression of cognitive decline. Despite enormous resources and efforts being invested, many therapeutic agents failed to prove the efficacy of the treatment of neurodegenerative dementia in clinical trials.

Presently cholinesterase inhibitor (ChEI) and memantine, an NMDA receptor antagonist, are the key pharmacological agents regularly employed in clinical settings. ChEI works by delaying the breakdown of acetylcholine in the brain, leading to increased acetylcholine concentrations in synaptic clefts and thus enhancing cholinergic neurotransmission [2]. Among them, donepezil was the first to receive approval for marketing in the United States back in 1996 and has since become the primary treatment for patients experiencing mild to moderate dementia. Recently the US Food and Drug Administration (FDA) has approved two breakthrough drugs, aducanumab [3] and lecanemab [4], which are disease-modifying therapies for Alzheimer’s disease (AD). Despite these advancements, challenges posed by their high cost, need for magnetic resonance imaging (MRI) monitoring, and intravenous administration, along with the importance of APOE genotyping, must be carefully addressed to ensure optimal patient care and treatment outcomes.

Being licensed and reimbursed by National Health Systems in most countries, ChEI drugs such as donepezil, galantamine, or rivastigmine are used as the standard treatment of neurodegenerative dementia. Clinical trials have shown that patients are modestly benefitted by ChEI with improved short-term and long-term outcomes in active daily living and neurocognitive functions [5, 6]. However, clinical practice reveals that only 30 to 60% of patients show modest clinical response [7]. In addition, adverse events of gastrointestinal symptoms may be intolerable that physicians must make decisions among drug dose tailoring, drug type switching, or treatment termination. It would be helpful to predict response prior to the treatment. Subsequent observational studies have identified various potential predictors of treatment response. Although not totally consistent, they include age, sex, education level, cognitive function at baseline, white matter hyperintensity, medial temporal atrophy, drug dose and type, and living condition [8]. However, gathering information as listed above is impractical. It remains physician’s empirical judgement to predict treatment response for each individual patient.

Mild cognitive impairment (MCI) is a diagnosis when patients show subtle signs of memory or other cognitive impairment without losing independent daily-living ability. MCI is not dementia, but patients are of high risk of converting to dementia, approximately 12% of annual conversion rate [9]. On the other hand, not all patients will convert to dementia, roughly 80% will convert and 20% will remain unchanged or revert to normal cognition [10]. Because of uncertain outcomes of MCI, treatment guidelines of MCI regarding the use of ChEI are inconsistent. Most guidelines do not recommend ChEI as a treatment option, but some leave the decision to clinicians’ discretion [11]. Recent trend of treatment strategy switches to more early stage of dementia, and MCI becomes the focus of research and development [12]. Currently, there are only a handful of papers reporting ChEI response in MCI, and the findings are inconsistent probably due to heterogeneity of this population [13, 14].

Brain age indicates the aging status of the brain which is estimated by a brain age estimation model. The model is trained by supervised machine learning algorithms based on neuroimaging data [15, 16]. The model building entails a training phase using a big neuroimaging dataset, called training set, from a group of cognitively and neurologically normal people and the tagged chronological age. The model is trained to predict an individual’s brain age based on the exact brain image of the individual. The neuroimaging dataset mostly uses brain MRI data, and different models could be trained based on different imaging substrates of the brain, resulting in gray matter brain age, white matter brain age or combined multimodal brain age. The gap between an individual’s predicted brain age and chronological age indicates the deviation from the aging trajectory of the normal people. The gap, named brain age gap or predicted age difference (PAD), has been found to be significantly increased in patients with neurological disease (such as AD, MCI, Parkinson’s disease, traumatic brain injury, epilepsy), psychiatric disease (such as schizophrenia, major depression), and metabolic disease (such as obesity and type II diabetes mellitus) [17 –24]. Interestingly, most diseases mentioned above are proved to be at risk of dementia.

Recently we used the data from the OASIS-3 databank, a longitudinal dataset from 1,098 people with dementia or normal cognition [25] and showed that PAD derived from white matter brain age can differentiate the severity of dementia as measured by the Clinical Dementia Rating global scores (CDR) [26] from 0 (normal cognition), 0.5 (very mild dementia), to 1 (mild dementia) [27]. Moreover, for patients with baseline CDR of 0.5, PAD can predict cognitive outcomes approximately 2 years later from the baseline. Specifically, patients who converted to dementia (CDR > 0.5) in 2 years exhibited PAD of 9.21 years, significantly larger than those whose CDR remained unchanged (PAD of 5.48 years) and those whose CDR reverted to normal cognition (PAD of 2.99 years). Similar results were also reported by Gaser et al. using gray matter brain age to predict AD conversion from MCI within 3 years of follow-up [28]. PAD has also been shown to be predictive of dementia progression in cognitively normal old people and in memory clinic patients [29, 30].

The aim of this study was to validate that gray matter brain age can demonstrate the effectiveness of ChEI treatment in patients with the CDR of 0.5. In this study we used the CDR of 0.5 to define our patient population because it has been used as one of the diagnostic criteria for MCI and mild AD in clinical research [12]. To study the treatment effect of ChEI, we employed gray matter brain age because neurons of cholinergic pathways reside ubiquitously in gray matter of the brain including the thalamus, striatum, limbic system, and neocortex [31]. We hypothesized that gray matter brain age is associated with cognitive outcomes after 2 years of ChEI treatment in patients with the CDR of 0.5. We compared PAD between patients with stationary or improved cognition and those with cognitive decline.

METHODS

Data source

The study was a retrospective study using existing data in Shuang Ho Hospital (New Taipei City, Taiwan). Patient data were retrieved from the ChEI treatment database which contained demographic records, cognitive assessment scores and brain MRI data. Normal data included demographic records and brain MRI data and were retrieved from the health check-up database. To build a gray matter brain age model, brain MRI data in the health check-up database were pooled together with MRI data of cognitively normal subjects in the OASIS-3 databank. The study was approved by the Institutional Review Board of Shuang Ho Hospital (TMU-JIRB Approval No: N202212005).

Cognitive assessment

Cognitive function was assessed using the CDR and Mini-Mental State Examination (MMSE) scores [32]. The CDR was administered by certified physicians or nurses to patients and informants. The interview consisted of 6 domains including memory, orientation, judgment and problem solving, community affairs, home and hobbies, and personal care. Each domain was graded by a five-point scale system to indicate the functional impairment: none = 0, questionable = 0.5, mild = 1, moderate = 2, severe = 3. The scores over the 6 domains were converted to a global score by an algorithm. The CDR of 0, 0.5, 1, 2, and 3 corresponded to normal cognition, very mild, mild, moderate, and severe dementia, respectively.

The MMSE was administered to patients by clinical neuropsychiatrists. MMSE comprised 8 categories of cognitive tests. The 8 categories and respective maximal points were as follow: orientation to time (5 points), orientation to place (5 points), registration (3 points), attention and calculation (5 points), recall (3 points), language (2 points), repetition (1 points), and complex commands (6 points). The total score was 30 points, and higher scores indicated better cognitive function.

MRI acquisition

MRI data in Shuang Ho Hospital were acquired on three MRI systems (GE HealthCare, Chicago, IL, US), two 1.5-Tesla scanners (Signa HDxt and Optima MR360) and one 3-Tesla scanners (Discovery MR750). Images were acquired using a routine clinical protocol including T1-weighted Fluid Attenuated Inversion Recovery (T1 FLAIR), T2-weighted FLAIR (T2 FLAIR) and T2-weighted Fast Recovery Fast Spin Echo (T2 FRFSE) pulse sequences. The matrix size of 512 and slice thickness of 7 mm were used uniformly across the three pulse sequences and three MRI systems. Other imaging parameters such as repetition time (TR), echo time (TE), inversion time (TI), and field of view (FOV) varied individually to optimize the scanning time. The imaging parameters for the T1 FLAIR sequence used: imaging plane = axial, TR = 1786 − 3017 ms, TE = 7.82–26.81 ms, TI = 738 − 1058 ms, field-of-view (FOV) = 220 − 250 mm. For the T2 FLAIR sequence, the imaging parameters were: imaging plane = axial, TR = 8000 − 9000 ms, TE = 117 − 148 ms, TI = 2000 − 2470 ms, FOV = 230 − 250 mm. For the T2 FRFSE sequence, the parameters were: imaging plane = coronal, TR = 3626 − 6773 ms, effective TE = 96 − 158 ms, FOV = 220 − 250 mm. The information about the scanner models and imaging parameters of the pulse sequences for patients and healthy participants was described in detail in Supplementary Material 1.

T1-weighted MRI data retrieved from OASIS-3 were acquired from 3 scanners (Siemens, Erlangen, Germany), one 3-Tesla Biograph mMR PET/MR scanner and two 3-Tesla TIM Trio scanner. The imaging pulse sequence used the 3-dimensional magnetization prepared rapid gradient echo sequence (T1 MPRAGE). The imaging parameters were: TR = 2,400 ms, TE = 3.16 ms, TI = 1,000 ms, FOV = 256 mm×256 mm×176 mm, and matrix size = 256×256×176.

Criteria for eligible data

Patient data were retrieved from the ChEI treatment database if they fulfilled the following criteria: 1) Age ranged from 50 to 85 years; 2) ChEI treatment lasted for at least 2 years; 3) Cognitive assessment including the CDR and MMSE was given at least twice. The first must be given upon entry of ChEI treatment, and another must be performed at least 2 years from the first; 4) Cognitive outcomes as rated with the CDR and MMSE were not contradictory, i.e., the CDR improved but the MMSE worsened, or vice versa; 5) MRI was performed within 120 days of the first cognitive assessment; 6) Patients had no history of substance abuse, head trauma, neurological, psychiatric, or metabolic disease.

According to the cognitive assessment at two time points, patient data were divided into two groups, ChEI-responsive and ChEI-unresponsive. Patients were categorized as the ChEI-responsive group if their CDRs were 0.5 at baseline and showed improved or stationary performance on the CDR and MMSE in the second assessment with respect to the first. Patients categorized into the ChEI-unresponsive group were those whose CDRs were 0.5 at baseline and showed worsening on both or one of the CDR or MMSE in the second assessment than the first.

Normal data retrieved from the health check-up database fulfilled the following criteria: 1) age 50 to 85 years old upon scanning, and 2) cognitively normal without history of substance abuse, head trauma, neurological, psychiatric, or metabolic disease.

Gray matter brain age model building

T1 FLAIR data from the Shuang Ho health check-up database and T1 MPRAGE data from cognitively normal subjects (CDR = 0) in OASIS-3 were pooled to build a gray matter brain age model. The resulting brain age model was then applied to T1 FLAIR data of patients in the Shuang Ho ChEI treatment database to calculate PAD for each patient.

Gray matter brain age model building underwent two major steps, data preprocessing and model construction (Fig. 1). Data preprocessing was conducted on T1-weighted images from each healthy participant, specifically T1-FLAIR or T1-MPRAGE, to derive gray matter features for the input to the training algorithm. In practice, the Segment toolbox of SPM12 [33] was employed to segment both gray and white matter components from the T1-weighted images. These components were registered to the ICBM-152 template, aligning diverse brain shapes to the standard Montreal Neurological Institute (MNI) space. The registration process utilized in-house code employing Large Deformation Diffeomorphic Metric Mapping (LDDMM), a generalized 3D nonlinear registration algorithm [34, 35]. The output of this registration was a deformation map, establishing the correspondence between the native individual space and the MNI space. Jacobian determinants of the deformation map were then calculated, representing the ratio of local volume change relative to the reference frame in the MNI space. The voxel size of the determinant map was 4 mm³. Finally, the determinants within the gray matter region were extracted as model features, comprising a total of 28,469 scalar values. These procedures were applied to both T1-FLAIR (Shuang Ho data) and T1-MPRAGE (OASIS-3 data). Notably, due to differing slice thicknesses in the two datasets, T1-FLAIR underwent up-sampling, while T1-MPRAGE underwent down-sampling during the feature extraction process.

Fig. 1

Flow chart of gray matter brain age model building.

The determinants of gray matter and chronological age of each participant in the training dataset were used as the inputs to a machine learning algorithm. We used Gaussian Process Regression (GPR) to construct the gray matter brain age model. The brain age approach conventionally suffered from the age-related bias, i.e., the brain age is over-estimated in younger age and under-estimated in older age. Several literatures have been proposed to address this bias. In the paper, we used the leave-one-out cross-validation (LOOCV) approach, which was similar to our previous paper [27], to evaluate the age-related bias and sex-related bias of the GPR model. Let the training set comprise N subjects, the LOOCV method used the image data and chronological age of N-1 subjects to construct an interim GPR model, followed by estimating the interim brain age of the left one. This process looped over the entire training set, so that every subject was associated with (x,y), and the bias-correction model was established through $y ’ = ay + b$ where x was the chronological age, y the interim brain age, and y’ the corrected brain age. To estimate the coefficients a and b, we minimized the absolute value of correlation coefficient between y’-x and x. Note that the bias-correction model was constructed for each sex separately in order to account for the sex-related bias.

To test the performance of the model, the mean of the absolute difference between predicted brain age and chronological age (mean absolute error, MAE) and Pearson’s correlation coefficient were adopted as test metrics. Since the model was to be applied to patient data which were acquired by the same 3 MRI scanners as the health check-up data, the test was performed on the same training data of Shuang Ho health check-up database using the leave-one-out technique.

Data analysis

The resulting gray matter brain age model was applied to T1 FLAIR of each individual patient to calculate brain age, and PAD was calculated by subtracting chronological age from predicted brain age.

Propensity score matching was applied to the eligible patient data to match ChEI-responsive subjects with ChEI-unresponsive subjects based on their propensity scores which were estimated using logistic regression. The method involved calculating the propensity score for each subject, specifically, calculating the probability of being in the responsive group given the covariates of age, gender, MMSE at baseline and medial temporal lobe atrophy (MTA) scores, followed by finding the matched subjects from the responsive and unresponsive groups who had similar propensity scores [36].

To ensure the balance of the potential covariates of ChEI response, demographic data of the matched ChEI-responsive and ChEI-unresponsive groups were compared. The demographic data included age, sex, education level, MMSE at baseline, ChEI agent type and dose, treatment duration, and living condition [37]. In addition, two imaging biomarkers, Fazekas scale and MTA score, were evaluated on T2 FLAIR and T2 FRFSE, respectively, by one of co-authors W.Y.I.T. who had more than 30 years of experience in diagnostic radiology. Fazekas scale ranged from 0 to 3 indicating the severity of white matter hyperintensity [38], and MTA score ranged from 0 to 4 indicting the severity of hippocampal atrophy [39].

Statistics

Demographic data were compared using two-sample t-test if the item belonged to continuous variable and using Chi-square test if the item belonged to categorical variable.

Normal distribution of PAD in the two patient groups was tested by Shapiro Wilk Test and variance equality was checked by Levene Test. Two-sample t-test was then performed to test the difference in PAD between ChEI-responsive and ChEI-unresponsive groups, and statistically significance was considered if p-value < 0.05.

To clarify the relationships of brain age with cognitive function and imaging biomarkers, the association analyses of baseline PAD with baseline MMSE, MTA, and Fazekas scale scores were performed.

RESULTS

Data selection results

The flow chart of data selection in the Shuang Ho ChEI treatment database and the resulting sample size was shown in Fig. 2. In a total of 1,054 patient data, 912 data were excluded due to ineligible conditions as described in Fig. 2. The propensity score matching was then applied to the 142 eligible patient data, and consequently, 58 matched data were obtained for each of the responsive and unresponsive groups.

Fig. 2

Flow chart of data selection from the ChEI treatment database.

Demographics of selected patients

Demographic data of the two groups were listed in Table 1. There was no significant difference between the two groups in each of the demographic items and imaging biomarkers.

Table 1

Demographics of the ChEI-responsive and ChEI-unresponsive groups

	ChEI-responsive (n = 58)	ChEI-unresponsive (n = 58)	p
Age (y)	75.16±7.48	75.26±8.06	0.946^a
Gender (M/F)	35(60%) / 23(40%)	34(59%) / 24(41%)	0.850^b
Treatment duration (y)	2.61±0.70	2.48±0.43	0.230^a
MMSE at entry	19.02±4.16	18.86±4.63	0.850^a
Education (y)	6.10±4.50	6.48±4.50	0.651^a
MTA	1.29±0.99	1.49±1.01	0.288^a
FAZEKAS	1.71±0.96	1.50±0.92	0.238^a
Living Status (n)			0.156^b
Solitary	4	1
With Family	52	46
Nursing Home	0	2
Missing living status	2	9
Medicine Agent [n(Dose)]			0.203^b (Agent)
Rivastigmine	21 (7.22 mg/day)	21 (7.09 mg/day)	0.860^a (Dose)
Donepezil	37 (8.58 mg/day)	37 (8.36 mg/day)	0.690^a (Dose)
Scanner (n)			0.111^b
Discovery MR 750	28	19
Optima MR360	11	9
Signa HDxt	19	30

(p-value^a): Two-sample t-test; (p-value^b): Chi-square test; Note: Rivastigmine and Donepezil are described in number of patients (shown left to the parenthesis) and averaged daily dose (shown inside the parenthesis).

Gray matter brain age model

We retrieved 72 T1 FLAIR data from the Shuang Ho health check-up database (female: n = 34, age: mean = 63.0, S.D. = 7.23, minimal = 49.8, maximal = 76.6, median = 64.0 years) and were co-trained with 290 T1 MPRAGE data from OASIS-3 (female: n = 171, age: mean = 67.84, S.D. = 7.63, minimal = 49.6, maximal = 84.7, median = 68.5 years) to build a gray matter brain age model.

There was no age-related dependence between PAD and chronological age after bias correction (Pearson’s correlation r = 0.07, p = 0.5589). The leave-one-out method showed MAE of 6.73 years for the gray matter brain age model after co-train process. There was significant association between brain age and chronological age (Pearson’s correlation r = 0.69, p = 2.026 × 10^–11, and R² = 0.4765).

Comparison of PAD between ChEI-responsive and ChEI-unresponsive groups

Shapiro Wilk Test showed normal distributions of PAD in ChEI-responsive (statistic = 0.988, df = 58, p = 0.814) and ChEI-unresponsive (statistic = 0.987, df = 58, p = 0.799) groups. Levene Test showed that variance equality of the two patient groups could be assumed (statistic = 0.042, p = 0.839). The ChEI-responsive group showed a lower value of PAD (3.87±9.02 years) than the ChEI-unresponsive group (8.44±8.78 years) (Fig. 3). A two-sample t-test showed that the difference was statistically significant (p = 0.0067, Cohen’s d = –0.51).

Fig. 3

Comparison of PAD between the ChEI-responsive and ChEI-unresponsive groups. Box plot was graphed according to 5–95 percentile; the dots above and below are the outliers. **p < 0.01.

Baseline PAD was significantly associated with baseline MMSE (Pearson’s correlation r = –0.285, p = 0.002) and MTA score (Pearson’s correlation r = 0.236, p = 0.011; Spearman’s correlation r = 0.230, p = 0.013) (Fig. 4). No significant association was found between PAD and Fazekas scale (Pearson’s correlation r = 0.102, p = 0.277; Spearman’s correlation r = 0.080, p = 0.392).

Fig. 4

Associations of baseline PAD with baseline MMSE, MTA and Fazekas scale, presented with scatter plots and regression lines. *p < 0.05, **p < 0.01.

DISCUSSION

In this study we have shown that gray matter brain age can demonstrate the effectiveness of ChEI after 2 years of treatment in patients with CDR of 0.5. Patients with cognitive decline post treatment exhibited advanced baseline brain age (PAD = 8.44±8.78 years) as compared with those without cognitive decline (PAD = 3.87±9.02 years). Presently it is impossible to predict outcomes of ChEI treatment, our results suggest that gray matter brain age could be used as a reference of prognosis in the management of patients with MCI or mild dementia. Another contribution of this study is that all imaging data directly come from routine clinical examinations. Gray matter brain age can be built readily from the real world data without additional imaging time, favoring its clinical usability.

To the best of our knowledge, this is the first study using brain age technology to demonstrate the effectiveness of ChEI treatment with MCI, i.e. CDR = 0.5. Previous studies used gray matter or white matter brain age at baseline to predict cognitive outcomes of MCI in 1 to 3 years [27, 28]. In those studies, ChEI treatment or other interventions were not prescribed, and so the prediction was made on the natural course of MCI. They found that conversion of MCI to AD or worsening of the CDR was associated with increased brain age. In the present paper, we further found that even with ChEI treatment, cognitive decline was also associated with increased brain age at baseline. The results imply that patients with advanced brain age presents unfavorable gray matter conditions of the brain, which is susceptible to AD conversion or cognitive decline, and is also more difficult to be benefitted from ChEI treatment. This explanation is plausible because atrophic gray matter, as represented by advanced gray matter brain age, might mean depleted post-synaptic neurons of the cholinergic pathways. In this case, even though acetylcholine is elevated by means of ChEI, the effect is suboptimal due to the paucity of post-synaptic neurons [7].

Interestingly, we found that gray matter brain age was significantly associated with MTA score, but not Fazekas scale (Fig. 4). The results highlight the role of gray matter brain age; it is a proxy of hippocampal atrophy, not white matter degeneration induced by small vessel disease. Our findings are consistent with the study by Cheng et al. who recruited AD patients with the CDR ranging from 0.5 to 2. They found that hippocampal atrophy but not white-matter hyperintensity predicts the long-term cognitive response to ChEI [40].

Multiple studies have investigated ChEI treatment outcomes, and the studies are heterogeneous, varying across the disease stages, different domains of outcomes and different follow-up time scales [7]. The disease stage ranged from MCI, mild, moderate to severe AD. The outcome domain included cognitive domain as assessed by the neuropsychological tests, MMSE, CDR or the Alzheimer’s Disease Assessment Scale - Cognitive, and functional domain such as Instrumental Activities of Daily Living or Physical Self-Maintenance Scale. The time scales of follow up varied from 6 months (short term) to 2 years or more (long term). In the present study, we focused on long term cognitive outcomes in patients with the CDR of 0.5 (presumably MCI) because this population is clinically relevant but most understudied.

In the present study, we found that PADs in ChEI-responsive and ChEI-unresponsive were significantly different. Given that there are potential covariates of ChEI treatment response [37], we compared the demographic variables of the two patient groups as listed in Table 1. We confirmed that there was no significant difference across all demographic variables. In addition, brain age results may be influenced by MRI scanners, and so we further compared the data distribution of the three scanners. Again, we found no significant difference between the two groups. Therefore, it is less likely that the significant difference in PAD was driven by the potential covariates or scanner types.

It is of note that a brain age model which has been trained by a dataset from a particular MRI machine cannot be directly applied to the data from another machine. There will be a large error in brain age prediction if applied directly, and so a process of model transfer is needed [41]. In the present study, we used a co-train process to facilitate model transfer [42]. Because 72 T1 FLAIR data of normal subjects in Shuang Ho health check-up database are insufficient to build a brain age model, we had to combine them with 290 T1 MPRAGE data of normal subjects in OASIS-3 to build the model. Having built the co-trained model, we were able to estimate brain age for patients in the Shuang Ho ChEI treatment database in which only T1 FLAIR data, instead of T1 MPRAGE data, was available.

In this study we used T1-weighted MRI of normal subjects in the Shuang Ho health check-up database and OASIS-3; their mean age was approximately 65 years. Previous studies using T1-weighted images of healthy participants with similar age range reported MAEs of approximately 4 to 5 years [29 , 44], superior to our MAE, 6.73 years. The reason may be attributed in part to the fact that our training set combined T1 FLAIR and T1 MPRAGE images, and they were acquired with very different slice thickness (7 mm for T1 FLAIR and 1 mm for T1 MPRAGE). Furthermore, our training set was obtained from six different MRI systems; three systems were manufactured by Siemens (all 3-T scanners), and another three systems by GE HealthCare (one 3-T and two 1.5-T scanners). Imaging parameters and spatial resolution were different across scanners. The heterogeneity of the scanners and associated image quality would arguably compromise the performance of the model. Despite unsatisfying performance of the real-world data, our brain age model could still demonstrate the effectiveness of ChEI treatment in patients with MCI.

Since 2021 FDA has approved several new drugs, e.g., aducanumab and lecanemab, for patients with MCI or mild dementia, and there are a few more drugs such as donanemab and solanezumab undergoing phase three clinical trials [12]. These drugs belong to monoclonal antibody drugs targeting amyloid-β fibrils, and have shown a modest effect of slowing down the cognitive decline (∼30%). However, a fraction of patients exhibited severe adverse reactions including brain edema and micro-bleeding [45, 46]. Since the present study has shown the capability of relating brain age to cognitive outcomes in CDR of 0.5 following ChEI treatment, it might be useful in patient stratification and prognostication in clinical trials of novel therapeutics.

The present study has limitations. Although the study showed promising results, it should be noted that this is a retrospective study using existing database from a single medical center. A prospective study across multiple centers is required to validate the results. The present study used the CDR of 0.5 to represent MCI. Although it is true in most patients, it might also include patients with mild AD or moderate AD in the study cohort [47]. Our results indicate that advanced brain age is associated with cognitive decline post ChEI treatment in patients with MCI. However, the results do not answer exactly how much cognitive decline ChEI can slow down. To answer the question, a longitudinal study on two patient groups, one with and one without ChEI treatment, is required. Although the gray matter brain age model showed significant difference in PAD, a substantial overlap was noted between the ChEI-responsive and ChEI-unresponsive groups (Fig. 3a). Using the present data, our ROC curve analysis gave a modest score of performance, area under the curve was 0.644 (Supplementary File 2). To meet the clinical standard, ROC performance must be improved by training a larger data with uniform distribution across different MRI scanners. Lastly, we did not include genotyping data such as APOE and CYPs in the study. It has been suggested that variation of these genes may affect the treatment response of ChEI [48]. However, a recent review on pharmacogenetic reports did not find definite associations of APOE or CYPs with ChEI treatment response [49].

In conclusion, the gray matter brain age model built from the real world MRI data can serve as a potential biomarker related to cognitive outcomes 2 years after ChEI treatment in patients with MCI. It might aid patient stratification and prognostication in the development of novel therapeutics in MCI.

AUTHOR CONTRIBUTIONS

Wen-Yih Isaac Tseng (Conceptualization; Data curation; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Software; Supervision; Writing – original draft; Writing – review & editing); Yung-Chin Hsu (Conceptualization; Funding acquisition; Investigation; Methodology; Project administration; Software; Supervision; Validation; Writing – original draft; Writing – review & editing); Li-Kai Huang (Conceptualization; Data curation; Investigation); Chien-Tai Hong (Conceptualization; Investigation; Methodology); Yueh-Hsun Lu (Data curation; Investigation; Resources); Jia-Hung Chen (Conceptualization; Data curation; Resources; Supervision; Validation); Chin-Kun Fu (Software); Lung Chan(Conceptualization; Data curation; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Resources; Supervision; Validation; Writing – original draft; Writing – review & editing).

Footnotes

ACKNOWLEDGMENTS

The authors are grateful to Miss Wu for her assistance in data retrieving and archiving.

FUNDING

The funding source of this study comes from a grant under the industry-academia collaborative contract (A-111-095) between AcroViz, Inc. and Taipei Medical University.

CONFLICT OF INTEREST

All authors receive funding from AcroViz, Inc. to conduct the reported study. All authors except Chin-Kun Fu have a patent for brain age application used in the reported study. Wen-Yih Isaac Tseng, Yung-Chin Hsu and Chin-Kun Fu are employees of AcroViz, Inc., and other authors are employees of Shuang Ho Hospital, Taipei Medical University. Wen-Yih Isaac Tseng and Yung-Chin Hsu own stock equity in AcroViz, Inc.

DATA AVAILABILITY

The data supporting the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

The supplementary material is available in the electronic version of this article:.

References

Nichols

, Steinmetz

, Vollset

, Fukutaki

, Chalek

, Abd-Allah

, Abdoli

, Abualhasan

, Abu-Gharbieh

, Akram

, et al.(2022) Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: An analysis for the Global Burden of Disease Study 2019. Lancet Public Health 7, e105–e125.

Birks

(2006) Cholinesterase inhibitors for Alzheimer’s disease. Cochrane Database Syst Rev 2006, CD005593.

Cavazzoni P, U.S. Food & Drug Administration, FDA’s Decision to Approve New Treatment for Alzheimer’s CORRECTED PROOF Disease, https://www.fda.gov/drugs/our-perspective/fdasdecision-approve-new-treatment-alzheimers-disease. Last updated June 7, 2021, Accessed on August 8, 2023.

U.S. Food & Drug Administration, FDA Converts Novel Alzheimer’s Disease Treatment to Traditional Approval, https://www.fda.gov/news-events/press-announcements/fda-converts-novel-alzheimers-disease-treatment-traditional-approval.Last updated July 6, 2023, Accessed on August 8, 2023.

, Garcia-Ptacek

, Jönsson

, Wimo

, Nordström

, Eriksdotter

(2021) Long-term effects of cholinesterase inhibitorson cognitive decline and mortality. Neurology 96, e2220–2230.

Yang

Y-H

, Wu

M-N

, Chou

P-S

, Su

H-C

, Lin

S-H

, Sung

P-S

(2018) Longitudinal neuropsychological outcome in Taiwanese Alzheimerမs disease patients treated with medication. Curr Alzheimer Res 15, 474–481.

Pozzi

, Conti

, Appollonio

, Ferrarese

, Tremolizzo

(2022) Predictors of response to acetylcholinesterase inhibitors in dementia: A systematic review. Front Neurosci 16, 998224.

Wattmo

, Wallin

ÅK

, Londos

, Minthon

(2011) Long-term outcome and prediction models of activities of daily living in Alzheimer disease with cholinesterase inhibitor treatment. Alzheimer Dis Assoc Disord 25, 63–72.

Petersen

, Smith

, Waring

, Ivnik

, Tangalos

, Kokmen

(1999) Mild cognitive impairment: Clinical characterization and outcome. Arch Neurol 56, 303.

10.

Tábuas-Pereira

, Baldeiras

, Duro

, Santiago

, Ribeiro

, Leitão

, Oliveira

, Santana

(2016) Prognosis of early-onset vs. late-onset mild cognitive impairment: Comparison of conversion rates and its predictors. Geriatrics 1, 11.

11.

Petersen

, Lopez

, Armstrong

, Getchius

TSD

, Ganguli

, Gloss

, Gronseth

, Marson

, Pringsheim

, Day

, Sager

, Stevens

, Rae-Grant

(2018) Practice guideline update summary: Mild cognitive impairment: Report of the Guideline Development, Dissemination, and Implementation Subcommittee of the American Academy of Neurology. Neurology 90, 126–135.

12.

Lee

, Howard

, Schneider

(2022) The current landscape of prevention trials in dementia. Neurotherapeutics 19, 228–247.

13.

Han

, Besser

, Xiong

, Kukull

, Morris

(2019) Cholinesterase inhibitors may not benefit mild cognitive impairment and mild Alzheimer disease dementia. Alzheimer Dis Assoc Disord 33, 87–94.

14.

Kasper

, Bancher

, Eckert

, Förstl

, Frölich

, Hort

, Korczyn

, Kressig

, Levin

, Palomo

MSM

(2020) Management of mild cognitive impairment (MCI): The need for national and international guidelines. World J Biol Psychiatry 21, 579–594.

15.

Cole

, Franke

(2017) Predicting age using neuroimaging: Innovative brain ageing biomarkers. Trends Neurosci 40, 681–690.

16.

Mishra

, Beheshti

, Khanna

(2023) A review of neuroimaging-driven brain age estimation for identification of brain disorders and health conditions. IEEE Rev Biomed Eng 16, 371–385.

17.

Beheshti

, Maikusa

, Matsuda

(2018) The association between“Brain-Age Score” (BAS) and traditional neuropsychologicalscreening tools in Alzheimerမs disease. Brain Behav 8, e01020.

18.

Chen

C-L

, Shih

Y-C

, Liou

H-H

, Hsu

Y-C

, Lin

F-H

, Tseng

W-YI

(2019) Premature white matter aging in patients with right mesial temporal lobe epilepsy: A machine learning approach based on diffusion MRI data. Neuroimage Clin 24, 102033.

19.

Chen

C-L

, Hwang

, Tung

Y-H

, Yang

L-Y

, Hsu

Y-C

, Liu

, Lin

Y-T

, Hsieh

M-H

, Liu

C-C

, Chien

Y-L

, Hwu

, Tseng

W-YI

(2022) Detection of advanced brain aging in schizophrenia and its structural underpinning by using normative brain age metrics. Neuroimage Clin 34, 103003.

20.

Chen

C-L

, Kuo

M-C

, Wu

W-C

, Hsu

Y-C

, Wu

R-M

, Tseng

W-YI

(2022) Advanced brain aging in multiple system atrophy compared to Parkinsonမs disease. Neuroimage Clin 34, 102997.

21.

Cole

, Leech

, Sharp

, the Alzheimerမs Disease Neuroimaging Initiative (2015) Prediction of brain age suggests accelerated atrophy after traumatic brain injury. Ann Neurol 77, 571–581.

22.

Franke

, Gaser

, Manor

, Novak

(2013) Advanced BrainAGE in older adults with type 2 diabetes mellitus. Front Aging Neurosci 5, 90.

23.

Koutsouleris

, Davatzikos

, Borgwardt

, Gaser

, Bottlender

, Frodl

, Falkai

, Riecher-Rossler

, Moller

H-J

, Reiser

, Pantelis

, Meisenzahl

(2014) Accelerated brain aging in schizophrenia and beyond: A neuroanatomical marker of psychiatric disorders. Schizophr Bull 40, 1140–1153.

24.

Ronan

, Alexander-Bloch

, Wagstyl

, Farooqi

, Brayne

, Tyler

, Cam-CAN , Fletcher

(2016) Obesity associated with increased brain age from midlife. Neurobiol Aging 47, 63–70.

25.

LaMontagne

, Benzinger

TLs

, Morris

, Keefe

, Hornbeck

, Xiong

, Grant

, Hassenstab

, Moulder

, Vlassenko

, Raichle

, Cruchaga

, Marcus

(2019) OASIS-3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease. medRxiv, doi:https://doi.org/10.1101/2019.12.13.19014902 [Preprint]. Posted December 15, 2019.

26.

Hughes

, Berg

, Danziger

, Coben

, Martin

(1982) A new clinical scale for the staging of dementia. Br J Psychiatry 140, 566–572.

27.

Tseng

W-YI

, Hsu

Y-C

, Kao

T-W

(2022) Brain age difference at baseline predicts clinical dementia rating change in approximately two years. J Alzheimers Dis 86, 613–627 .

28.

Gaser

, Franke

, Klöppel

, Koutsouleris

, Sauer

, Alzheimerမs Disease Neuroimaging Initiative (2013) BrainAGE in mild cognitive impaired patients: Predicting the conversion to Alzheimerမs disease. PLoS One 8, e67346.

29.

Wang

, Knol

, Tiulpin

, Dubost

, De Bruijne

, Vernooij

, Adams

HHH

, Ikram

, Niessen

, Roshchupkin

(2019) Gray matter age prediction as a biomarker for risk of dementia. Proc Natl Acad Sci U S A 116, 21213–21218.

30.

Biondo

, Jewell

, Pritchard

, Aarsland

, Steves

, Mueller

, Cole

(2022) Brain-age is associated with progression to dementia in memory clinic patients. Neuroimage Clin 36, 103175.

31.

Hampel

, Mesulam

M-M

, Cuello

, Farlow

, Giacobini

, Grossberg

, Khachaturian

, Vergallo

, Cavedo

, Snyder

, Khachaturian

(2018) The cholinergic system in the pathophysiology and treatment of Alzheimerမs disease. Brain 141, 1917–1933.

32.

Folstein

, Folstein

, McHugh

(1975) “Mini-mental state”. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 12, 189–198.

33.

Ashburner

, Friston

(2005) Unified segmentation. Neuroimage 26, 839–851.

34.

Beg

, Miller

, Trouvé

, Younes

(2005) Computing large deformation metric mappings via geodesic flows of diffeomorphisms. Int J Comput Vis 61, 139–157.

35.

Miller

, Trouvé

, Younes

(2006) Geodesic shooting for computational anatomy. J Math Imaging Vis 24, 209–228.

36.

Rosenbaum

, Rubin

(1983) The central role of the propensity score in observational studies for causal effects. Biometrika 70, 41–55.

37.

Wattmo

(2013) Prediction models for assessing long-term outcome in Alzheimerမs disease: A review. Am J Alzheimers Dis Dementiasr 28, 440–449.

38.

Kim

, MacFall

, Payne

(2008) Classification of white matter lesions on magnetic resonance imaging in elderly persons. Biol Psychiatry 64, 273–280.

39.

Fazekas

, Chawluk

, Alavi

, Hurtig

, Zimmerman

(1987) MR signal abnormalities at 1.5 T in Alzheimerမs dementia and normal aging. Am J Roentgenol 149, 351–356.

40.

Cheng

Y-W

, Chen

T-F

, Cheng

T-W

, Lai

Y-M

, Hua

M-S

, Chen

Y-F

, Chiu

M-J

(2015) Hippocampal atrophy but not white-matter changes predicts the long-term cognitive response to cholinesterase inhibitors in Alzheimerမs disease. Alzheimers Res Ther 7, 72.

41.

Jiang

, Guo

, Du

, Xu

, Qiu

(2019) Transfer learning on T1-weighted images for brain age estimation. Math Biosci Eng 16, 4382–4398.

42.

Chen

C-L

, Hsu

Y-C

, Yang

L-Y

, Tung

Y-H

, Luo

W-B

, Liu

C-M

, Hwang

T-J

, Hwu

H-G

, Isaac

Tseng W-Y

(2020) Generalization of diffusion magnetic resonance imaging–based brain age prediction model through transfer learning. Neuroimage 217, 116831.

43.

Cole

(2020) Multimodality neuroimaging brain-age in UK biobank: Relationship to biomedical, lifestyle, and cognitive factors. Neurobiol Aging 92, 34–42.

44.

Franke

, Gaser

(2019) Ten years of BrainAGE as a neuroimaging biomarker of brain aging: What insights have we gained? Front Neurol 10, 789.

45.

Salloway

, Chalkias

, Barkhof

, Burkett

, Barakos

, Purcell

, Suhy

, Forrestal

, Tian

, Umans

, Wang

, Singhal

, Budd Haeberlein

, Smirnakis

(2022) Amyloid-related imaging abnormalities in 2 phase 3 studies evaluating aducanumab in patients with early Alzheimer disease. JAMA Neurol 79, 13.

46.

Filippi

, Cecchetti

, Spinelli

, Vezzulli

, Falini

, Agosta

(2022) Amyloid-related imaging abnormalities andβ-amyloid–targeting antibodies: A systematic review. JAMA Neurol 79, 291.

47.

Wada-Isoe

, Kikuchi

, Umeda-Kameyama

, Mori

, Akishita

, Nakamura

, ABC Dementia Scale Research Group (2019) Global clinical dementia rating score of 0.5 may not be an accurate criterion to identify individuals with mild cognitive impairment. J Alzheimers Dis Rep 3, 233–239.

48.

Cacabelos

(2020) Pharmacogenetic considerations when prescribing cholinesterase inhibitors for the treatment of Alzheimerမs disease. Expert Opin Drug Metab Toxicol 16, 673–701.

49.

Zúñiga Santamaría

, Yescas Gómez

, Fricke Galindo

, González González

, Ortega Vázquez

, López López

(2022) Pharmacogenetic studies in Alzheimer disease. Neurologia (Engl Ed) 37, 287–303.