Machine learning prediction of future amyloid beta positivity in amyloid-negative individuals

Background The pathophysiology of Alzheimer’s disease (AD) involves \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β-amyloid (A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β) accumulation. Early identification of individuals with abnormal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β-amyloid levels is crucial, but A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β quantification with positron emission tomography (PET) and cerebrospinal fluid (CSF) is invasive and expensive. Methods We propose a machine learning framework using standard non-invasive (MRI, demographics, APOE, neuropsychology) measures to predict future A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β-positivity in A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β-negative individuals. We separately study A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β-positivity defined by PET and CSF. Results Cross-validated AUC for 4-year A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β conversion prediction was 0.78 for the CSF-based and 0.68 for the PET-based A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β definitions. Although not trained for the clinical status-change prediction, the CSF-based model excelled in predicting future mild cognitive impairment (MCI)/dementia conversion in cognitively normal/MCI individuals (AUCs, respectively, 0.76 and 0.89 with a separate dataset). Conclusion Standard measures have potential in detecting future A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β-positivity and assessing conversion risk, even in cognitively normal individuals. The CSF-based definition led to better predictions than the PET-based definition. Supplementary Information The online version contains supplementary material available at 10.1186/s13195-024-01415-w.


Introduction
Alzheimer's disease (AD) is a common neurodegenerative disease with a complex and unclear pathway and a long prodromal phase.The progressive and irreversible nature of AD highlights the need for detecting early changes in the brain that occur decades before dementia.Research on amyloid, tau, and neurodegeneration (ATN) biomarkers, in accordance with the NIA-AA 2018 framework [1], has focused on tracking AD progression through beta-amyloid (Aβ ) and Tau protein accumulations.These biomarkers are associated with neurodegeneration and cognitive decline [2].Identifying amyloid burden in cognitively normal individuals holds promise for identifying those at risk of developing AD [3], and it is expected to become the standard for prescribing A β-targeted drugs [4].Currently, cerebrospinal fluid (CSF) and positron emission tomography (PET) with 18F-labeled amyloid tracers are established methods for confirming the presence and measuring the extent of A β accumula- tion in the brain [5].PET can detect metabolic and biochemical alterations in the brain deviating from normal.The clearance efficiency of A β protein aggregates can be detected in CSF [6].CSF peptides (Aβ 1-42 ) and hyperphosphorylated tau are correlated with amyloid plaques and neuronal tangles observed in brain autopsies [7] and are linked to cognitive decline, offering valuable insights into early detection.Despite their significance, PET and CSF are not widely available.CSF can cause discomfort and are invasive, while PET involves exposure to radiation and requires specialized equipment and personnel.Additionally, PET and CSF results can disagree because they measure different aspects of amyloid pathology [8,9].Therefore, promoting the use of universally available data, such as demographics and cognitive scores, is important to advance AD research [10,11].
Standardized and longitudinal datasets such as ADNI (Alzheimer's Disease Neuroimaging Initiative), provide valuable resources for developing machine learning (ML) models with different feature combinations to study PET and CSF biomarkers.These ML models can highlight the important modalities for identifying at-risk individuals transitioning from normal control (NC) to mild cognitive impairment (MCI) and also potentially to dementia by tracking changes in A β-positivity states.However, there are limited studies addressing the future predictability of A β-positivity using widely available measures [12].Addi- tionally, the comparative analysis of model performance based on categorizing individuals as A β-positive or A β -negative using either CSF or PET biomarkers remains understudied.
This study aims to develop an ML-based approach to predict the conversion to A β-positivity in individu- als who are A β-negative using data available widely in clinical settings.We utilize demographic data (age, gender, education), APOE4 (genetic), neuropsychological scores, and MRI-derived brain volumes to predict whether an individual with A β-negative will convert to A β-positive (referred to as A β-Converter) or remain A β -negative (referred to as A β-Stable) over a 4-year period.Analyses are performed separately and in parallel for grouping individuals based on CSF and PET biomarkers.We analyze the role of different data types in the predictive performance of the model in each CSF and PETbased cohort.Additionally, we examine whether baseline CSF/PET measures improve the predictive performance of the model and how that affects the contribution of cognitive measures and MRI biomarkers.We also investigate the role of baseline PET measures in predicting A β -positivity based on CSF (Aβ42) and vice versa, the role of baseline CSF measures in predicting A β-positivity based on PET.Finally, we evaluate our model's predictive capability for conversion to MCI/dementia in healthy and MCI individuals.Our study extends to demonstrate that both classification and regression approaches display similar trends.

ADNI data
Data used in this work were obtained from the ADNI (http:// adni.loni.usc.edu).The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD.The primary goal has been to test whether serial MRI, PET, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD.For up-to-date information, see http:// www.adni-info.org.
This study included participants from all phases of ADNI with baseline demographics, APOE4, psychological test results, and MRI biomarkers, who also had available longitudinal CSF measures or 18 F-florbetapir (AV45) PET measures.

CSF and PET measures
CSF samples were collected and processed as previously described [13].The concentration of A β42, pTau, and tTau in CSF was measured using the fully automated Elecsys immunoassays (Roche Diagnostics, Basel, Switzerland) by the ADNI biomarker core (University of Pennsylvania, Philadelphia, PA).We obtained these measures from the ADNI depository (UPENNBI-OMK9_04_19_17.csv, UPENNBIOMK10_07_29_19.csv,UPENNBIOMK12_01_04_21.csv).Several cutoffs have been proposed and evaluated for determining A β-positivity based on CSF A β-42 measure, such as 980 pg/mL determined through ROC analyses with FBP PET as the endpoint [14].We selected a cutoff 880 pg/mL, which was derived based on the BioFINDER study, accounting for pre-analytic differences and consequently validated in ADNI [15].We selected this cutoff for two distinct reasons: (1) among suggested cutoffs it maximizes the number of participants, which is important due to limited participant numbers, particularly in the A β-Converter group; (2) the cutoff has been developed independently of ADNI data but validated on ADNI, which we consider as important characteristic [15,16].
Omitting the FreeSurfer QC had no major effect on the predictive accuracy, but increased the number of samples.A similar conclusion was previously reached by Gómez-Sancho et al. [24] indicating that whether or not ADNI FreeSurfer segmentations were quality controlled was immaterial to MRI-based MCI-to-dementia conversion prediction.

Study cohorts
To develop our classification models for predicting progression toward A β-positivity, we assigned participants into two overlapping cohorts based on the availability of longitudinal CSF measures or 18 F-florbetapir (AV45) PET measures.Specifically, we compiled a CSF-based cohort and a PET-based cohort.The subject selection procedure for each cohort is visualized in Fig. 1.The labeling of participants was constrained by the availability of data and the established cut-off values.
For the CSF-based cohort, we selected participants with available baseline CSF data (N = 1415).Among these participants, we first selected the initially A β-negative individuals (N = 726) and then we excluded participants (1) with an initial diagnosis of dementia (N = 35), (2) without available follow-up CSF sample (N = 345), (3) with a follow-up duration of fewer than 4 years without conversion to A β-positivity (N = 172), and (4) with an unstable A β+/Aβ-status during the follow up (N = 12).Finally, among the remaining 162 individuals, we defined A β-negative individuals who converted positive within the available follow-up time period as A β-Converter (N = 52) and individuals who remained A β-negative for more than 4 years as A β-Stable (N = 110).In order to balance the dataset, we selected a set of auxiliary data from the initially A β-positive group.This step was crucial because the number of participants in the A β-Converter group was considerably lower than in the A β-Stable group, which could negatively impact the classification performance.Specifically, we selected individuals from the initially A β-positive group (N = 689) whose A β 42 values were close to the cutoff used for classifying participants into the A β-positive and A β-negative groups, i.e., individuals with A β42 > 680 pg/mL and A β42 < 880 pg/mL (N = 244).The cut-off points of 680 and 880 were chosen for auxiliary data selection in such a way that a reasonable number of samples were obtained to balance the data set.We further excluded 47 individuals with an initial dementia diagnosis, 102 individuals without a follow-up CSF sample, and 17 individuals who returned to A β-negative status during the available follow-up.The remaining participants (N =78) are used as A β-Converters, but only during the training phase for balancing the dataset.This approach helped to create a more balanced dataset, which was necessary to develop a robust classification model.
The same procedure was applied for selecting the PET-based cohort, which resulted in a dataset of 53 individuals as A β-Converters and 178 individuals as A β -Stables.To create an auxiliary dataset from the initially A β-positive group, cut-off points of 1.11 and 1.25 were used.As a result, 82 individuals were included in the auxiliary dataset for use as A β-Converters only during the training phase.The selection of these cut-off points was done to ensure a reasonable number of samples were obtained to balance the dataset.
It is important to emphasize that participants who are A β-positive at baseline, with their A β values close to the cutoff point, i.e., 78 participants in CSF-cohort and 82 participants in PET-cohort, are employed only during the training of the model and serve as A β -Converters for balancing the dataset during the training.They are not used in the testing of the model.
A follow-up period of 4 years was chosen for the A β -Stable group in order to exclude participants with shorter follow-up duration.This decision was made to prevent potential false negatives while ensuring a reasonable number of participants within the A β-Stable group.Increasing the follow-up period beyond 4 years would have led to a marked reduction in the number of participants.
To develop our regression models for predicting future A β42, we only selected subjects with available baseline and longitudinal CSF data.Among those, we eliminated participants with dementia diagnosis at baseline and then we restricted our selection to participants with CSF follow-up samples for at least 4 years.If an individual had multiple follow-up samples after 4 years, the first sample (closest to 4 years) was used as the future sample for the analysis.Two hundred fifty-three individuals were included in our dataset as a result.Similarly, we selected  For independent validation of our models, we selected a fully non-overlapping cohort from the one used for training.The independent validation was done to evaluate our CSF-based and PET-based A β-positivity conver- sion prediction models for MCI/dementia conversion prediction in healthy and MCI individuals.Since A β plays a significant role in the initiation of the AD process, the A β-positivity conversion prediction model should be able to detect AD-related cognitive decline even though it is not specifically developed to predict conversion to MCI/dementia.To this end, we selected two different datasets for assessing each CSF/PET-based model: The first dataset was for evaluating the future conversion of MCI/dementia in CN individuals and the second dataset was for the evaluation of future conversion of dementia in MCI individuals.For the evaluation of the CSF-based model, we considered all ADNI participants with a baseline diagnosis of CN or MCI who were not utilized for training the CSF-based approach and who had available baseline demographics, APOE4, neuropsychological test results, and MRI biomarkers.We determined two groups based on baseline and longitudinal diagnosis labels, regardless of CSF/PET biomarker status: (a) Stable group: CN/MCI individuals who have remained stable for 5 years or more following baseline, (b) Converter group: (1) Individuals with CN diagnosis at baseline who later develop MCI or dementia (the last two diagnoses must be MCI or dementia), and (2) MCI individuals who later develop to dementia (the last two diagnoses must be dementia).This resulted in a dataset of 67 converters and 131 stables for conversion prediction in CN individuals and a dataset of 211 converters and 126 stables in conversion prediction in MCI individuals (Table 2).The same procedure was used to select data for the PET-based model evaluation, resulting in a dataset of 74 converters and 95 stables for conversion prediction in CN individuals and a dataset of 234 converters and 90 stables in conversion prediction in MCI individuals (Table 3).

Machine learning framework
We developed a classifier based on the ridge logistic regression (RLR) approach [26] to predict the conversion of A β-positivity within 4 years in A β-negative individu- als in both the CSF-based cohort and PET-based cohort.The framework of the classification procedure is shown in Fig. 2.
The number of participants in our PET-cohort and CSF-cohort differs because individuals in each cohort were selected based on the availability of CSF/PET data (Fig. 1).To ensure a fair comparison between the CSF and PET experiments, we decided to include an equal number of participants for CSF-based and PET-based Table 1 Characteristics of the study cohorts: Age, education, A β42, and global SUVR measures are reported as mean(standard deviation).CN: cognitively normal, SMC: subjective memory concern (participants with self-reported significant memory concern), MCI: mild cognitive impairment, EMCI: early MCI, LMCI: late MCI.Classification of EMCI and LMCI is done by ADNI based on the WMS-R Logical Memory II Story A score.The specific cutoff scores were as follows (out of a maximum score of 25): EMCI was assigned for a score of 9-11 for 16 or more years of education, a score of 5-9 for 8-15 years of education, or a score of 3-6 for 0-7 years of education.LMIC was assigned for a score of ≤ 8 for 16 or more years of education, a score of ≤ 4 for 8-15 years of education, or a score of ≤ 2 for 0-7 years of education [25]

Aβ-Converter Aβ-Stable Auxiliary data Regressioncohort
Sample size (N) model training.Given that the PET-based dataset contains more individuals, we randomly subsampled it to achieve the same number of participants as in the CSFbased model for the training phase.
We designed three models with different feature combinations for each PET-based and CSF-based A β-positivity prediction task: The first model was trained using only demographic data and APOE4, the second model was trained using neuropsychological test results in addition to demographic data and APOE4, and the third model was trained using all demographic data, APOE4, neuropsychological test results, and MRI biomarkers.
In addition to the classification task, we developed a regression strategy based on ridge linear regression [27] to predict future A β 42 (CSF-based) and global SUVR (PET-based) values from multimodal data similar to the   1, the participants in the CSF-cohort originated from different ADNI cohorts, including ADNI1.The participants in ADNI1 underwent 1.5-Tesla (1.5-T) T1-weighted MRI images, whereas participants in other ADNI cohorts underwent 3T MRI scans.To account for differences in MRI biomarkers caused by different field strengths, we applied ComBat [28], as a domain adaptation method to the MRI biomarkers before performing the actual machine learning procedure.ComBat employs an empirical Bayes method for batch correction in microarray expression data, and it has also been successfully utilized for domain adaptation in imaging data [29][30][31].We then compared the results obtained with and without applying ComBat.However, the addition of a domain adaptation step using Com-Bat approach did not affect the results.Therefore, we excluded it and applied our machine learning approach to the original MRI biomarkers.
The Supplementary Table S 1 summarizes the experiments reported in different subsections of the " Results" section.The codes for classification and regression analyses are available at "https:// github.com/ Elahe hMora di/ AB-posit ivity-predi ction".

Implementation and performance evaluation
We used two nested cross-validation loops (10-fold for each loop) to evaluate the model's performance and estimate the model's parameter ( ) to prevent overfitting and maximize performance.First, an external 10-fold crossvalidation was implemented in which samples were randomly divided into 10 subsets with the same proportion of each class label (stratified cross-validation).At each step, a single subset was left for testing and the remaining subsets were used for training.Again, the training set was divided into 10 subsets used to select the model's parameter ( ).The optimal parameters were selected in classification according to the misclassification error and in regression according to MSE( the mean square error) across the 10-fold of the inner loop.The performance of the model was then evaluated based on mainly AUC (area under the receiver operating characteristic curve) in classification analyses and correlation score in regression analyses in the test subset of the outer loop.We also provided balanced accuracy, sensitivity, and specificity for the classification analyses and mean absolute error for regression tasks in Supplementary materials.The reported results in the " Results" section are averages over 10 nested 10-fold CV runs.Repeated CV was used to reduce variability due to the partitioning of the data.

Predicting PET and CSF A β-positivity in A β-negative individuals from multimodal data excluding PET and CSF baseline measures
We predicted the progression to A β-positivity based on CSF and PET data.As explained in the "Methods" section, we designed three models with different feature combinations for each classification of PET-based and CSF-based future A β-positivity prediction as well as for the regression model of predicting future A β 42 and global SUVR values.Figure 3 shows the results of all these computational analyses.These results are the average over 10 times repeated 10-fold cross-validation analyses for each method.
According to the findings (Fig. 3a, b, c, d; Fig. S1), CSF-based models outperformed PET-based models in both classification and regression tasks.The average AUC value across 10 computation runs was 0.78 with a 95% confidence interval (95% CI) of 0.70 to 0.85 with all demographic, APOE4, neuropsychological test results, and MRI biomarkers for predicting CSF-based progression to A β-positivity and 0.61 (95% CI of 0.53 to 0.70) for predicting PET-based progression to A β-positivity.The average correlation score for predicting future A β 42 was 0.45 (95% CI of 0.37 to 0.53), and the average correlation score for predicting future global SUVR was 0.40 (95% CI of 0.31 to 0.49), with all demographic, APOE4, neuropsychological test results, and MRI biomarkers.
We explored the characteristics of individuals not correctly classified by the A β-positivity prediction model, particularly regarding their A β-values to determine if misclassified cases are closer to the positivity cutoff.For CSF-based experiments, we used the model with all measures, including demographics, APOE4, neuropsychological test results, and MRI, as it demonstrated superior performance.In PET-based experiments, the model with only demographics and APOE4 measures was chosen, exhibiting the best performance in the PET-based model.Figure S 3 presents a box plot of the A β-value (Aβ 42 for CSF and global SUVR for PET) for the last visit, utilized for labeling in each group.However, no significant differences in A β-values were observed between correctly classified and misclassified individuals in both CSF-based and PET-based experiments.We also examined their diagnosis at baseline; however, there were no significant differences observed between correctly classified and misclassified individuals.
The addition of neuropsychological test results and MRI biomarkers to the demographics and APOE4 data increased the performance of the CSF-based A β-positivity conversion prediction model as shown in Fig. 3a.The average AUC value for predicting A β-positivity in A β-negative individuals increased from 0.72 (95% CI of 0.63 to 0.80) to 0.78 (95% CI of 0.70 to 0.85), although the improvement was not statistically significant (p = 0.18) according to Delong's test.The average correlation value for predicting future a β 42 value also increased slightly, from 0.44 (95% CI of 0.35 to 0.52) to 0.45 (p = 0.50).However, adding neuropsychological test results and MRI biomarkers to demographic and APOE4 measures did not improve the performance of the PET-based models.Our results showed that including these variables decreased the performance of the PET-based A β-positivity conversion prediction model.The average AUC value decreased from 0.68 (95% CI of 0.60 to 0.76) to 0.61 (95% CI of 0.53 to 0.70).The average correlation value for predicting future global SUVR value also slightly decreased, from 0.41 (95% CI of 0.32 to 0.49) to 0.40 (95% CI of 0.31 to 0.49).
We investigated the contribution of different variables in CSF-based and PET-based models, accounting for all demographics, APOE4, neuropsychology test results, and MRI biomarkers (Fig. 3e and f, Fig. S2). Figure 3e and f show the coefficient values for 100 RLR classification models, derived from 10 runs of 10-fold CV, with a single column heatmap representing the correlation score between each variable and the label (Aβ-Stable, A β-Converter), and a bar graph indicating the significance of each variable, calculated by taking the mean of the absolute values of the regression coefficients.
In PET-based models, the primary contributors were APOE4 and age (as shown in Fig. 3f ), whereas in CSFbased models (Fig. 3e), several variables from different data types are significantly contributing, with APOE4 being the most significant, followed by the volume of ventricles and TRABSCORE (Trail Making Test Part B).Additionally, the correlation panels in Fig. 3e and f, Tables S 2, and S 3 show a stronger correlation between various neuropsychological test results and MRI biomarkers with CSF-based labeling compared to PETbased labeling.This explains the higher contribution of different variables in CSF-based modeling, as well as the superior performance of CSF-based modeling.
The difference in the performance of the PET-based and CSF-based A β positivity is quite interesting.How- ever, it is important to note that CSF and PET measures capture distinct aspects of the underlying biology, and, of course, the choice of CSF measure for defining A β posi- tivity significantly influences the results.We investigated the correlation between A β 42 and the A β42/Aβ 40 ratio with PET global SUVR using ADNI data.The correlation between A β 42 and PET global SUVR was − 0.59, whereas the correlation between the A β42/Aβ 40 ratio and PET global SUVR measure was − 0.73 (Fig. S4).However, we defined A β-positivity based on A β 42 alone, since the A β 40 measure was only available for a small number of participants.The dataset with the A β42/Aβ 40 ratio is rather limited, comprising 13 A β-Converters and 63 A β -Stable cases, with the converter group being particularly underrepresented.Due to the small size, we are concerned that it may not be sufficient for the development of a robust predictive model, thus putting the reliability of the results at risk.

Predicting PET and CSF A β-positivity in A β-negative individuals from multimodal data including PET and CSF baseline measures
We investigated the use of baseline CSF measures, i.e., A β 42, pTau, and Tau measures, to predict future CSF-based A β-positivity and the use of baseline PET measures, i.e., global and regional SUVR measures, to predict future PET-based A β-positivity.Our goal was to assess the validity of baseline CSF and PET measures for predicting future conversion.To gain insight into the contribution of other data types besides CSF baseline and PET baseline measures, we designed four models with different feature combinations.We further developed a regression-based approach to predict the changes (future -baseline) in the A β 42 and global SUVR values.We predicted the changes in A β42/global SUVR values rather than their future val- ues because of the high correlation between the baseline and future values of these measures.Figure 4 and Fig. S5 show the results of all these computational analyses.
To enhance the robustness and reliability of our results, we chose to exclude individuals with atypical changes: a significant decrease in PET-based A β values and a notable increase in CSF-based A β values with increas- ing age.Individuals who showed a marked reduction in PET-based A β values and a pronounced rise in CSF- based A β values over a 4-year period were considered atypical.Such cases could potentially indicate issues with data acquisition.Specifically, we removed 11 individuals from the CSF regression cohort with an increase in A β 42 higher than 500 pg/mL, and similarly, 11 individuals were removed from the PET regression cohort with a decrease in global SUVR greater than 0.1.Although the number of participants with such changes in A β values was relatively small, their removal had a clear impact on the model performance of the CSF-based model but not of the PET-based model.In the CSF-based model, excluding outliers significantly improved the model's effectiveness.Specifically, in predicting changes in A β 42 over 4 years using CSF measures, the correlation score improved from 0.21 to 0.44 upon removing outliers.Conversely, in PET-based analyses for predicting changes in global SUVR based on baseline PET measures, the correlation score slightly decreased after outlier exclusion, from 0.30 to 0.27.The outlier removal was clearly more important for the CSF-based model probably due to the existence of obvious outliers in CSF values but not in PET values.
As expected, using baseline CSF measures to predict future CSF-based A β-positivity and using baseline PET measures for predicting future PET-based A β-positivity resulted in relatively high performance.Specifically, for CSF-based A β-positivity prediction, the average AUC significantly increased to 0.93 when using only baseline CSF measures, compared to an AUC of 0.78, the highest achieved in experiments without CSF baseline data.Similarly, in the context of PET-based A β-positivity pre- diction, the average AUC increased to 0.78 with only baseline PET measures from 0.68, the best performance without baseline PET data (as detailed in the " Predicting PET and CSF A β-positivity in A β-negative individuals from multimodal data excluding PET and CSF baseline measures" section).
However, CSF-based prediction outperformed PETbased prediction in terms of performance (as shown in Fig. 4a).The average AUC value over 10 computation runs was 0.93 (95% CI of 0.88 to 0.97) for CSF-based prediction, whereas the average AUC value was 0.78 (95% CI of 0.71 to 0.84) for PET-based prediction.Moreover, the average correlation value for predicting the changes in A β 42 was 0.44 (95% CI of 0.27 to 0.56), whereas for predicting the changes in global SUVR value, the average correlation score was 0.27 (95% CI of 0.17 to 0.35).Interestingly, adding other data types did not improve the  3 performance of classification and regression tasks in both CSF-based and PET-based analyses (Fig. 4a, Fig. S5).
We explored the impact of various variables, including demographics, APOE4, neuropsychological test results, and MRI biomarkers, with baseline CSF/PET measures in both PET-based and CSF-based models.Figure 4d and e visualizes the coefficient values for the top 20 predictors.Although the addition of additional data types did not improve overall model performance, the visualization of coefficient values emphasized substantial contributions from specific features, particularly APOE4, when integrated into the model.Particularly in the CSF-based model (Fig. 4d), other features from neuropsychological test results and MRI measures also exhibited significant contributions.Despite the absence of performance improvement, these findings emphasize the importance of specific features beyond baseline CSF/PET measures.

Predicting CSF-based future A β-positivity from CSF and PET baseline measures
We further extended our analysis by predicting CSFbased future A β-positivity based on baseline CSF meas- ures and baseline PET measures.We selected only individuals with available PET baseline measures in the CSF-based cohort, resulting in a dataset of 84 A β-Stables, 36 A β-Converters, and 50 A β-positive as auxiliary data (for using as A β-Converters only in the training phase to balance dataset).We developed a classification model to predict future A β-positivity with three different feature combinations: the first model was trained based on CSF measures only, the second model based on PET measures only, and the third model based on both CSF and PET measures.The results are shown in Fig. 5a and Fig. S6.As anticipated, using baseline CSF measures for predicting CSF-based future A β-positivity resulted in improved performance compared to using baseline PET measures.Interestingly, the addition of baseline PET measures with baseline CSF measures did not improve the performance of the classification model for predicting future A β-positivity.The average AUC was 0.93 (95% CI of 0.87 to 0.98) using only CSF measures, 0.82 (95% CI of 0.72 to 0.90) using only PET measures, and 0.91 (95% CI of 0.85 to 0.96) using both CSF and PET measures.
To predict changes in A β42, first, we selected 169 individuals from the CSF regression-cohort who had available PET baseline measures.We then removed six individuals as outliers, with the changes in A β 42 higher than 500.Again we designed the regression model for predicting the changes in A β 42 using three different fea- ture combinations: CSF measures only, PET measures only, and the combination of PET and CSF measures.The resulting average correlation value were 0.41 (95% CI of 0.23 to 0.55) using only CSF measures, 0.12 (95% CI of 0.004 to 0.25) using only PET measures, and 0.36 (95% CI of 0.20 to 0.50) using both CSF and PET measures.Similar to our classification results, the best predictive performance was achieved using only baseline CSF measures.S7.Surprisingly, the use of baseline CSF measures to predict future PET-based A β-positivity resulted in improved performance compared with the use of baseline PET measures, the average AUC value increased from 0.79 (95% CI of 0.72 to 0.85) to 0.85 (95% CI of 0.78 to 0.91) (p-value = 0.18).In addition, a combination of CSF and PET measures provided similar performance to the use of CSF measures alone, with an average AUC value of 0.83 (95% CI of 0.76 to 0.88).The p-value was 0.049 between the model based on PET measures alone and the model with the combination of PET and CSF measures.

Predicting PET-based
To predict changes in global SUVR, first, we selected 353 individuals from the PET regression-cohort who had available CSF baseline measures.We then removed 10 individuals as outliers, with the changes in global SUVR less than − 0.1.Finally, the remaining 343 individuals were used for predicting the changes in global SUVR with different feature combinations.The results are shown in Fig. 5b and Fig. S7.Unlike the classification results, the average correlation value was rather near to one another in all three different feature combinations.The average correlation value was 0.30 (95% CI of 0.22 to 0.39) with CSF measures alone, 0.32 (95% CI of 0.22 to 0.42) with PET measures alone, and 0.33 (95% CI of 0.22 to 0.43) with CSF and PET measures combined.
We further investigated the contribution of different CSF and PET measures in both classification and regression models.Figure 5d and e show the coefficient values of the 20 variables with the highest importance calculated by taking the mean of the absolute values of the coefficients.Notably, A β 42 was the most important vari- able in both classification and regression models (Fig. 5d  and e).Furthermore, the two important CSF measures, i.e., A β 42 and TAU, rank among the top ten most impor- tant variables.These findings demonstrate the importance of CSF measures in predicting PET-based future A β-positivity and also provide an explanation of why the classification model based on baseline CSF measures outperformed the model based on baseline PET measures.
It is essential to emphasize that, in utilizing baseline PET measures, we incorporated all available regional SUVR values, even though some may not provide significant information.In contrast, CSF consists of only three main features.This distinction may influence the significance of the contribution of PET global SUVR to the modeling process compared to the contribution of CSF A β 42 measure.To assess the importance of the baseline global SUVR measure in predicting PET-based A β-positivity, we devised two distinct classification models.The first exclusively featured PET global SUVR and CSF A β 42 as predictors, while the second included all CSF measures (Aβ42, PTAU, and TAU) alongside five PET baseline SUVR measures-specifically, global SUVR (SUMMA-RYSUVR_WHOLECEREBNORM) and four additional SUVR measures used in creating the global SUVR (FRONTAL_SUVR, CINGULATE_SUVR, PARIETAL_ SUVR, TEMPORAL_SUVR).The results are shown in Fig. S8.The AUC of the first model with CSF A β 42 and PET global SUVR as predictors was 0.81, and the AUC for all three CSF measures and five PET SUVR measures was 0.82.This indicates a slight decrease compared to using all regional SUVR and CSF measures (AUC: 0.83).Although AUC performance was similar across all three models, an examination of sensitivity revealed that the model with only two predictors exhibited relatively low sensitivity and balanced accuracy (Fig. S8c).However, the model with five SUVRs and CSF measures demonstrated performance comparable to the model incorporating CSF measures with all SUVR measures.Investigating the contribution of CSF A β 42 and PET global SUVR measures, when only five SUVRs are included, demonstrates that the importance of A β 42 and global SUVR is nearly the same (Fig. S8a).
Additionally, we compared the performance of a model with only five SUVRs against a model with all regional SUVR measures.The AUC was 0.79 when considering all baseline regional SUVRs, and 0.73 when using only five SUVRs.These findings suggest that if CSF baseline measures are not available during the learning process, the inclusion of all SUVR measures improves the performance.Given the application of regularized logistic regression, we anticipate the model to adeptly select the most significant information from all available predictors, even when all regional SUVRs are employed (Fig. S8c).

MCI/dementia conversion prediction
We validated the relevance of our A β-positivity predic- tion models, both CSF-based and PET-based models, by applying them to an independent dataset (Tables 2 and 3) to predict changes in clinical status from cognitively normal to MCI/dementia or from MCI to dementia.The predictors were demographics, APOE4, neuropsychological test results and MRI biomarkers (the model trained in the " Predicting PET and CSF A β-positivity in A β-negative individuals from multimodal data excluding PET and CSF baseline measures" section).
We trained a model for predicting CSF-based future A β-positivity using all available participants (Fig. 1a, Table 1) and used it to classify CN-Stable vs. CN-Converter groups, as well as MCI-Stable vs. MCI-Converter groups.We next repeated the process for the PET-based model.The results are shown in Fig. 6.The CSF-based model performed quite well, achieving an AUC of 0.76 (95% CI of 0.68 to 0.83) for MCI/dementia conversion prediction in the CN group and an AUC of 0.89 (95% CI of 0.86 to 0.93) for dementia conversion prediction in the MCI group.However, the PET-based model had lower performance, with an AUC of 0.60 (95% CI of 0.51 to 0.69) for classifying CN-Stables vs. CN-Converters and an AUC of 0.72 (95% CI of 0.67 to 0.78) for classifying MCI-Stables vs. MCI-Converters.
Given that neither the CSF-based nor the PET-based model was specifically designed for classifying CN-Stables vs. CN-Converters/MCI-stables vs. MCI-converters, and that both models used basic and noninvasive data (demographics, APOE4, cognitive, and MRI), the CSFbased model performed exceptionally well for the MCI/ dementia conversion prediction in CN and MCI groups, while the PET-based model performed well only in the MCI group.However, in the CN group, the PET-based model failed to predict conversion to MCI/dementia.

Discussion
Determining A β-status is crucial for the prescription of amyloid-targeted treatments in the future.This study aimed to predict conversion to A β-positivity in A β -negative individuals using data that is widely available in clinical settings.To achieve this goal, we developed a classification model based on an RLR approach incorporating demographics, APOE4, neuropsychological tests, and MRI biomarkers.We categorized participants into A β-positive and A β-negative groups, based on CSF A β 42 and PET global SUVR.Due to categorization inconsistencies [41], we conducted separate analyses with participants classified as A β+/Aβ -based on CSF and PET biomarkers.
The selection of a suitable cutoff point for categorizing subjects into A β-positive and A β-negative groups is important in the prediction of A β-positivity in A β-negative individuals as it can impact the results of our analysis.Particularly, in CSF-based grouping, the selection of the cutoff point is challenging due to the presence of various cutoff values in the existing literature [15,16].We decided to use the cutoff 880 pg/mL, a value established from predictions made in the BioFINDER study, independently from the ADNI dataset.The selection of a cutoff point determined in independent data is statistically appropriate.Additionally, our choice of the 880 pg/ mL cutoff, as opposed to larger cutoff values, allowed us to include a larger number of participants in the A β-Converter group.However, to confirm the reliability of our classification results and to eliminate any potential bias arising from the cutoff point for categorizing participants into A β+/Aβ− groups, we conducted regression analy- ses that were aligned with our classification experiments for predicting future CSF A β (Aβ42) or PET A β (global SUVR) values via a ridge linear regression.These steps collectively contribute to the robustness and credibility of our study's findings.
In our analyses, we demonstrated that features from multiple modalities, including demographics, neurophysiological scores, APOE4, and MRI biomarkers, can model the progression to A β-positivity detected by either CSF or PET biomarkers.Interestingly, our findings indicate that utilizing CSF A β 42 for participant categoriza- tion resulted in more accurate predictions of future A β -positivity compared to PET global SUVR.Also, the CSFbased model benefited from additional features, whereas the accuracy of the PET-based model decreased with more features.In more detail, the AUC of CSF-based model increased from 0.72 with APOE and demographics as the features to 0.78 with all the features.However, the AUC of the PET-based decreased from 0.68 with APOE4 and demographics as the features to 0.61 to all the features.To comprehend the performance disparity between the PET-based and CSF-based models, we investigated the contributions of different data types in each model.We observed that cognitive scores and MRI biomarkers, along with APOE4 as the most important variable, strongly contribute to the CSF-based model.In contrast, in the PET-based model, the contribution of cognitive and MRI biomarkers was smaller, and the primary predictors were APOE4 and age.
It is essential to highlight that CSF A β 42 and PET global SUVR measures, utilized in defining A β-positivity, employ different mechanisms for detecting A β protein.
PET imaging reveals the presence of amyloid plaques in the brain, whereas CSF analysis is associated with the clearance of amyloid from the brain.Low CSF amyloid levels may indicate inefficient clearance, leading to brain amyloid accumulation [42,43].Although we used the same ML framework for predicting future A β-positivity, direct comparison of the results is challenging.However, our findings suggest that predicting future CSF-based A β -positivity is relatively easier compared to PET-based A β -positivity.In a previous study, Jagust and Landau [44] explored the factors influencing the transition from A β -negative to A β-positive, as determined by PET meas- ures.Their findings indicated that age, baseline PET, and being a female APOE4 carrier were associated with an increased risk of conversion.Interestingly, in another study conducted by Elman et al. [45], cognitive tests were found to predict conversion to A β-positivity, which was defined based on CSF and/or PET.The discrepancy between these findings can be understood based on the results of the current study, where prediction performance improved by incorporating cognitive measures for CSF-based cohort but not in the PET-based cohort.
Determining A β-positivity using CSF biomarker can be done either based on A β 42 levels and the A β42/Aβ 40 ratio.In this study, we defined A β-positivity solely based on A β42, as the A β 40 measure was only available for a small number of participants.However, recent studies have emphasized the relevance of the A β42/Aβ 40 ratio in defining A β-positivity using CSF biomarker [46], as it exhibits a stronger correlation with PET-based A β meas- ure compared to A β 42 alone.Moreover, the correlation between the ratio of A β42/Aβ 40 and global SUVR is higher than the correlation between A β 42 and the global SUVR measure in ADNI data.We speculate that utilizing the A β42/Aβ 40 ratio for predicting CSF-based A β -positivity may yield results that align more closely with PET-based predictions.Nevertheless, due to the limited number of available samples, we cannot conduct experiments to confirm this hypothesis.
Previous studies have primarily focused on detecting A β-positivity at the time of the study [47][48][49][50][51] (baseline prediction).These studies have employed either machine learning algorithms or statistical analyses to identify A β-positivity at baseline and explore the relationships between cognitive measures, various biomarkers, and A β -positivity.For example, Palmqvist et al. [47] developed a model utilizing demographics, cognitive tests, white matter lesions, APOE, and plasma biomarkers (Aβ42/Aβ40, tau, and neurofilament light chain) to detect A β-positivity at baseline.Their model achieved an AUC of 0.80-0.82when trained on BIOFINDER dataset [48,52] and validated on ADNI data.Similarly, another study by the same group [48], detected baseline A β-positivity using demo- graphic, APOE, and cognitive information and achieved an AUC of 0.65 in cognitively healthy individuals.However, we are aware of only two studies that have focused on predicting future A β-positivity [12,45].The first study, by Elman et al. [45], examined the association of baseline cognitive measures with progression to A β-positivity in A β-negative individuals.The second study, by Park et al. [12], employed machine learning algorithms to predict future A β-positivity in A β-negative individuals.In Park et al. [12] study, they used a PET biomarker for subject classification in A β+/Aβ -groups and developed a classifier with baseline age, gender, APOE4 genotype, and PET SUVR measures in ADNI data.They achieved a cross-validated AUC of 0.67 using basic demographic and genetic factors, which improved to 0.84 when including PET SUVR measures.Our work has a similar objective to the study by Park et al. [12], but we focused on widely available non-invasive measures to predict future A β conversion.Moreover, a key characteristic of our study was to assess the future predictability of A β-positivity determined based on either CSF and PET biomarkers whereas Park et al. [12] only considered a PET-based definition.
We validated the relevance of our prediction models for future A β-positivity in an independent dataset com- posed of separate ADNI participants, predicting clinical status changes from CN to MCI/dementia or from MCI to dementia.The CSF-based model performed well, achieving an AUC of 0.76 for MCI/dementia conversion prediction in CN individuals and an AUC of 0.89 for the dementia conversion prediction in MCI individuals.However, the PET-based model performed worse in predicting conversion, reaching an AUC of 0.60 for CN to MCI/dementia conversion and an AUC of 0.72 for MCI to dementia conversion.The CSF-based model's excellent performance in predicting MCI/dementia conversion in a completely independent dataset is intriguing, indicating its potential for detecting AD/dementia-related changes at an early stage.Although the model was not specifically designed to classify CN/MCI-Converter versus CN/ MCI-Stable groups, the CSF-based model demonstrated strong performance comparable to existing studies designed for such classification tasks [53][54][55].
We analyzed the use of baseline CSF and PET measures for predicting future A β-positivity.Including baseline scores significantly improved the performance of the prediction models.Notably, when baseline CSF/PET measures were available, the addition of other data types such as cognitive scores and MRI biomarkers contributed very little value.Moreover, utilizing baseline CSF/PET measures enabled the prediction of the change/rate in A β 42 and global SUVR values which are rather hard to model [56].However, the invasive nature, high cost, and limited availability of CSF/PET measures restrict their usage.An alternative approach could be to predict future A β-positivity using predicted CSF and PET measures derived from other variables.This strategy could offer more robust insights, particularly when substituting less accessible measures with more readily available or cost-effective options.A promising area for future research lies in exploring the prediction of PET and CSF values from blood measures, which are generally more accessible.
Moreover, we investigated the role of baseline CSF and PET measures in both predicting CSF-based future A β -positivity as well as in predicting PET-based future A β -positivity.Interestingly, baseline CSF measures showed superior predictive ability for future PET-based A β-positivity compared to baseline PET measures.A closer look at the coefficient values in the PET-based model revealed that A β 42 made the greatest contribution to the model.Recent studies suggest that AD-related alterations are detectable earlier in CSF than in PET [57,58], which may explain the superior performance of baseline CSF measures in predicting PET-based A β-positivity.
Our study has several limitations.First, the sample sizes were small due to our specific inclusion criteria, although the data came from one of the largest longitudinal dementia-prediction cohorts (ADNI).To compensate for the small sample size, we evaluated our CSF-based/ PET-based models in an independent dataset for MCI/ dementia conversion prediction.Second, the sizes of A β -Stable and A β-Converter groups were different, which we partially addressed by introducing auxiliary data for label balancing during model training.Third, the CSFbased and PET-based analyses used different datasets, challenging direct performance comparison.To address this, we included an equal number of participants for CSF-based and PET-based model training to facilitate a fairer comparison.Fourth, we utilized a single ML approach to combine various data types and develop our models.While this approach worked well for CSF-based analyses, a more advanced approach may be needed for PET-based future A β-positivity prediction.Last, the models to predict future A β-positivity in A β-negative individuals were not evaluated on an independent database.While the validation of the prediction models in an independent dataset composed of separate ADNI participants, predicting clinical status changes from CN to MCI/dementia or from MCI to dementia, provides a partial remedy, it does not completely overcome the limitation.

Conclusion
We developed ML-based models to predict future A β -positivity in A individuals, determined based on either CSF or PET biomarkers.The CSF-derived dichotomization achieved better predictive performance (AUC = 0.78) compared to PET dichotomization (AUC = 0.68).The discrepancy in performance may be attributed to the different mechanisms of A β detection and the dis- cordance between the two biomarkers.Further research is required to understand this discrepancy.However, by using non-invasive measures including demographics, APOE4, neuropsychological scores, and MRI biomarkers, our CSF-based A β-positivity conversion prediction model performed well in identifying A β-Stables vs A β -Converters with an AUC 0f 0.78, as well as in identifying CN-Converters vs. CN-Stables (AUC = 0.76) and MCI-Stables vs. MCI-Converters (AUC = 0.89).These findings demonstrate the significance of neuropsychological and MRI biomarkers to detect the risk of conversion to AD, even in cognitively normal individuals.The detection may occur before current thresholds for A β-positivity are reached, providing an opportunity for early intervention.

Fig. 1
Fig. 1 Data selection procedure for predicting progression to A β-positivity for a CSF-cohort and b PET-cohort.The data from Stable A+ individuals, who are A β-positive at baseline with their A β values close to the cutoff point, are employed only during the training of the model and serve as A β -Converters for balancing the dataset during the training.They are not used in the testing of the model

Fig. 3
Fig. 3 Predicting future A β-positivity from multimodal data excluding PET and CSF baseline measures: a Bar plots showing the average AUC and average correlation score across 10 computation runs for CSF-based and PET-based models, with 95% confidence intervals error bars.b Distribution of probability score derived by RLR for PET-based and CSF-based prediction in A β-Stable and A β-Converter groups.c, d Scatter plot for estimation of A β 42 (c) and global SUVR (d) derived by ridge linear regression (with demographics, APOE4, neuropsychology, and MRI biomarkers).The results in b, c, and d are from 1 computation run with median performance.e, f Heatmap of coefficient values across 10 runs of 10-fold CV (100 models) for CSF-based models (e) and pet-based models (f), with a single column heatmap representing the correlation score between each variable and the label (Aβ-Stable, A β-Converter), and a bar graph showing the importance of each predictor calculated by the mean of the absolute value of regression coefficients derived by RLR.There are 100 columns in the heatmaps, with each column representing the coefficient values for one model.The performance of predicting A β-positivity in A β-negative individuals was higher with CSF-cohort compared to PET-cohort, suggesting the higher relevance of CSF data for conversion prediction.ADAS13:Alzheimer Disease Assessment Scale, 13 items, ADASQ4: ADAS Delayed Word Recall, MMSE: Mini-Mental State Examination score, RAVLT: Rey's Auditory Verbal Learning Test, LDELTOTAL: Logical Memory Delayed Recall Total, TRABSCOR: Trail Making Test Part B, FAQ: Functional Assessment Questionnaire, CDRSB: Clinical Dementia Rating-Sum of Boxes, ICV: intracranial volume, MidTemp: middle temporal gyrus

Fig. 4
Fig. 4 Predicting future A β-positivity from multimodal data including PET and CSF baseline measures: a Bar plots showing the average AUC and average correlation score (predicting the difference between future and baseline A β42/global SUVR measures) across 10 computation runs for CSF-based and PET-based models, with 95% confidence intervals error bars, CSF/PET stands for CSF baseline measures (Aβ42, pTau, Tau) for predicting CSF-based A β-positivity and PET measures (global and regional) for predicting PET-based A β-positivity.b, c Scatter plot for estimation of the difference between future and baseline A β 42 (c) and the difference between future and baseline global SUVR (d) derived by ridge linear regression (with CSF/PET measures).The results are from 1 computation run with median performance.d, e Heatmap of coefficient values across 10 runs of 10-fold CV (100 models) for CSF-based classification model (d) and PET-based regression model (e), with the bar graphs showing the importance of each predictor calculated by the mean of the absolute value of regression coefficient.The results of experiments without PET and CSF baseline measures are shown in Fig. 3 future A β-positivity from CSF and PET baseline measures We continued our investigation by predicting PET-based future A β-positivity on baseline CSF and PET measures.Again, we limited the PET-based dataset to individuals with available CSF baseline measures, resulting in a dataset of 161 A β-Stables and 45 A β-Converters, and 72 A β -positive as auxiliary data (for using as A β-Converters only in the training phase to balance dataset).We developed a classification model to predict future PET-based A β-positivity with three different feature combinations: The first model was trained based on CSF measures only, the second model was based on PET measures only, and the third model was based on both CSF and PET measures.The results are shown in Fig. 5b-e and Fig.

Fig. 5
Fig. 5 Predicting future CSF-based and PET-based A β-positivity from baseline CSF and PET measures: a Bar plots showing the average AUC and average correlation score across 10 computation runs for predicting future CSF-based amyloid positivity, with 95% confidence intervals error bars.b Bar plots showing the average AUC and average correlation score across 10 computation runs for predicting future PET-based amyloid positivity, with 95% confidence intervals error bars.c Distribution of probability score derived by RLR for PET-based prediction in A β-Stable and A β -Converter groups with CSF and PET baseline measures.d, e Heatmap of coefficient values across 10 runs of 10-fold CV (100 models) for PET-based classification model (d) and pet-based regression model (e), with the bar graphs showing the importance of each predictor calculated by the mean of the absolute value of regression coefficient.The regression model is designed for predicting the difference between future and baseline global SUVR measures

Fig. 6
Fig. 6 MCI/dementia conversion prediction in CN and MCI groups: a Boxplot probability score derived by RLR for CSF-based and PET-based model for classification of stables vs converters in CN and MCI groups.b ROC curves of CN and MCI subjects classification to stable and converter groups using CSF-based and PET-based models, with AUC values and 95% confidence intervals Table 1 and participant's RIDs are available as Supplementary material.

Table 2
Characteristics of the validation cohorts for performance validation of CSF-based model.Age and education are reported as mean(standard deviation)

Table 3
Characteristics of the validation cohorts for performance validation of PET-based model.Age and education are reported as mean(standard deviation) Schematic representation of the classification framework classification task with three models using various feature combinations.By applying a regression model, we were able to eliminate the effects of the cutoff point for classifying participants into A β+/Aβ -groups on the performance.Moreover, as presented in Table