Optimizing timing and cost-effective use of plasma biomarkers in Alzheimer’s disease

Abstract

Background and objectives

Early and cost-effective identification of amyloid positivity is crucial for Alzheimer’s disease (AD) diagnosis. While amyloid PET is the gold standard, plasma biomarkers such as phosphorylated tau 217 (pTau217) provide a potential alternative. This study evaluates the diagnostic accuracy of a combined-panel approach using machine learning models and evaluated the biomarker significance.

Methods

We enrolled 371 participants, including AD (n = 143), non-AD (n = 159), and cognitively unimpaired (CU, n = 69) controls. Combined panels of pTau217, pTau181, glial fibrillary acidic protein (GFAP), neurofilament light chain (NFL), Aβ42/40, and total tau were measured prior to the amyloid PET scan. The multiclass logistic (LR) regression, support vector machines, decision trees, and random forests (RF)—were applied to classify amyloid positivity (A+) at all stages or at early clinical stages (1–3). In AD, we tested whether the biomarker may define the clinical stagings.

Results

When benchmarked against amyloid PET, plasma biomarker–based stratification achieves an optimal balance between diagnostic accuracy and cost-effectiveness. The multi-class LR performed equivalently with RF model in identifying A+. The combined plasma panel reached an > 92% accuracy in identifying A+, with performance increasing to 93.4% at early clinical stages. We ranked the importance of individual biomarkers and pTau217 alone achieved comparable accuracy (> 90%) and was the top-ranked biomarker in the LR or RF model. NFL and GFAP correlated significantly with Mini-Mental State Examination; however, these plasma biomarkers did not enhance clinical staging stratification.

Discussion

The use of multiclass LR model enhances amyloid classification, particularly at earlier clinical stages. While the combined-panel approach is most accurate, pTau217 alone provides a cost-effective alternative for screening. These findings support the integration of plasma biomarkers and ML into clinical workflows for early detection and patient stratification.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-025-01851-2.

Introduction

In 2024, plasma biomarkers were incorporated into the updated criteria for diagnosing Alzheimer’s disease (AD) [1]. This revision classified imaging and fluid biomarkers, each serving a complementary but non-interchangeable role. Among these biomarkers, plasma phosphorylated tau 217 (pTau217) has shown promise in clinical applications for predicting amyloid positivity and tau burden [2–6]. Other plasma biomarkers, such as Aβ42/40, pTau181, neurofilament light chain (NFL), and glial fibrillary acidic protein (GFAP), have been classified as core or disease-staging markers [7–9]. Among plasma biomarkers, pTau217 has demonstrated the strongest associations with amyloid PET positivity and has been shown to differentiate AD from non-AD conditions earlier and with greater specificity than pTau181 and Aβ42/40 [4]. The FDA has granted traditional approval to two disease-modifying agents for AD, with more favorable treatment outcomes observed at earlier clinical stages. A key issue in applying plasma biomarkers is determining whether a single biomarker can achieve both diagnostic and clinical staging accuracy comparable to a combined biomarker panel [1].

In statistical modeling, multi-class logistic regression (LR) allows the conversion of outcomes into categorical variables while incorporating multiple independent variables. Recent advances in machine learning (ML) algorithms have gained interest as alternative methods for classification problems. These strategies include support vector machines (SVM) [10–12], decision trees [13], and random forests [14, 15]. Given the high-dimensional nature of biomarker data and the complex, non-linear interactions among predictors, ML models may offer a promising approach to enhance diagnostic accuracy and efficiency compared to conventional regression methods [16–18]. However, the reliability of risk predictions is often not explicitly included in ML models, and overfitting remains a concern when assessing performance improvements [19]. Head-to-head comparisons between multi-class LR and ML models can provide deeper insights into real-world data. Additionally, applying feature selection in ML models may significantly improve clinical understanding for patient selection.

Given the high cost and limited accessibility of amyloid PET imaging, particularly in low-resource settings, a single biomarker is preferable to a full panel, as it significantly reduces costs and simplifies interpretation. While plasma biomarkers offer a less invasive alternative to PET, the timing of blood draw relative to imaging and the incremental cost of multi-marker panels critically determine their clinical utility. This study evaluates not only diagnostic performance but also when and how a minimal biomarker set maximizes cost-effectiveness in real-world practice. In this context, this study aimed to compare the performance of a combined plasma biomarker panel (pTau217, pTau181, NFL, GFAP, total tau, and Aβ42/40) and to rank the importance of individual biomarkers from three clinical perspectives:

1. Diagnosis of amyloid positivity in a heterogenous population- including cognitively unimpaired (CU) controls and a spectrum of neurodegenerative conditions—was used to assess whether individual plasma biomarkers could approximate the diagnostic accuracy of the full composite panel in identifying amyloid-positive AD, thereby mirroring the real-world challenge faced in memory-clinic settings.

2. Early-Stage Detection – We refined the analysis to include only CU individuals and patients in the early disease stages (clinical stages 1–3 [1]) and evaluated the same diagnostic scenario for AD detection.

3. Clinical Staging Stratification in Amyloid-Positive AD – Among AD patients with confirmed amyloid positivity, we compared the performance of a single plasma biomarker and the full biomarker panel in predicting early (stage 1–3) versus late (stage 4–6) clinical stages [1].

Methods

Patient enrolment

This prospective study was conducted by the Declaration of Helsinki and was approved by the Institutional Review Board of Kaohsiung Chang Gung Memorial Hospital. Individuals were prospectively enrolled from the Kaohsiung Chang Gung Memorial Hospital [20, 21] for comprehensive clinical and imaging evaluations.

CU was enrolled according to the following inclusion criteria: (1) age 45 years; (2) not having cognitive symptoms as assessed by a behavioral neurologist using their judgment of symptoms on the National Alzheimer’s Coordinating Center B9 form; (3) not fulfilling the criteria for mild or major neurocognitive disorders according to the Diagnostic and Statistical Manual of Mental Disorders, fifth edition (DSM-V); and (4) a clinical dementia rating global score 0.

The inclusion criteria for the mild cognitive impairment (MCI) patients were: (1) age over 45 years; (2) memory complaints corroborated by an informant; (3) objective memory impairment adjusted for age and education, as judged by a physician; (4) preserved general cognitive functioning, with a clinical dementia rating (CDR) score of 0.5 and Mini-Mental State Examination (MMSE) score of 24–30; and (5) not fulfilling the criteria for dementia according to the DSM-V.

The inclusion criteria for the dementia patients were: (1) age over 45 years; (2) major cognitive symptoms with significant impairment of daily life activities; (3) CDR score ≥ 0.5 and MMSE score < 24; and (4) fulfilling the dementia criteria according to the DSM-V.

The exclusion criteria for the study were a history of clinical stroke during the screening phase, a modified Hachinski ischemic score of ≥4, clinically unmanaged diabetes, major depressive disorder, or dysthymic disorder.

Clinical evaluation

After the screening, all individuals underwent inclusion and exclusion criteria checkups, and blood samples were collected < 6 months prior to amyloid PET imaging. The demographic data in this study included apolipoprotein E4 (ApoE4) status, years of education, age at onset and gender. Apolipoprotein E4 (APOE4) genotype was determined using single nucleotide polymorphisms rs7412 and rs429358 [22]. The obtained ApoE4 genotypes were dichotomized into ε4 carriers (heterozygous or homozygous) and non-carriers (i.e., ε2 or ε3 carriers). Age was calculated by subtracting the age at plasma sampling and birthday. At the time of plasma biomarker collection, all participants underwent the Mini-Mental State Examination (MMSE), clinical dementia rating (CDR) and CDR sum of the box (CDR-SOB). Clinical staging evaluations were performed independently of the aforementioned cognitive tests or the biological staging system. The clinical staging was based on established criteria [1], which define six clinical stages ranging from stage 1 (asymptomatic) to stage 6 (severe dementia). In this study, stages 1 to 3 were categorized as early-stage, while stages 4 to 6 were considered late-stage.

Biological diagnosis stratifications

The AD continuum classification was based on the 2024 NIA-AA criteria [1]. Amyloid positivity was first determined by two experienced nuclear medicine specialists. This was further supported by amyloid centiloid quantitative cutoff values (> 26.8) by optimizing the agreements between visual reads and amyloid centiloid derived from the Bi-Gaussian Mixture Model distribution in the Taiwan-ADNI cohort (supplementary Fig. 1A). Among the 143 cases showing amyloid positivity, 128 patients underwent [18 F] Florzolotau PET imaging, all exhibiting positive visual readouts, and 126 showed positive quantitative results (Supplementary Fig. 1B). The supplementary material method section discusses imaging acquisition, processing and quantification.

A total of 371 cases completed the study and were stratified into three biological categories: cognitively unimpaired (CU, n = 69, A-T-), AD continuum (n = 143, A + T+), and non-AD continuum (A-T- or A-T+, n = 159). For study purposes, we also classified the groups into amyloid positive (A+, n = 143) and amyloid negative groups (A-, n = 228).

We included amyloid-negative neurodegenerative disorders as part of the non-AD continuum (n = 159) to facilitate the mixture of heterogeneity. The non-AD continuum was composed of three subgroups, defined according to the following criteria.

1. Amyloid- and Tau-Negative neurodegenerative Disorders (A-T-): This subgroup included patients with Parkinson’s disease dementia (n = 30) [23], ), Lewy body dementia (n = 35) [24], and multiple system atrophy (n = 6) [25].

2. Tau-First Cognitive Proteinopathy (TCP, n = 60): TCP was defined by PET criteria showing A-T + status and clinical phenotypes characterized by non-executive dysfunctions [21].

3. Progressive Supranuclear Palsy (PSP) Spectrum (n = 28): PSP cases exhibited positive [18 F]Florzolotau PET uptake in the brainstem and subcortical nuclei, consistent with the characteristic distribution of tau pathology [26]. None of the PSP cases demonstrated amyloid positivity.

Although both the TCP and PSP-spectrum groups share an A–T + biological profile, they differ markedly in clinical presentation and regional tau pathology. TCP cases show early cortical tau deposition accompanied by non‐executive cognitive impairment, whereas PSP‐spectrum cases exhibit predominant brainstem and subcortical tau uptake alongside characteristic executive and oculomotor deficits.

Blood sample collection and analysis

Fasting venous blood samples were obtained using EDTA as the anticoagulant at baseline. The supernatant plasma was aliquoted into 1 cc polypropylene tubes and frozen at -80 °C. All plasma and DNA samples were analyzed at the Linkou Chang Gung Memorial Hospital Laboratory.

Single-molecule array analysis was used for the ultrasensitive quantification of AD biomarkers (HD-X instrument, Quanterix, USA). Plasma Aβ1-42, Aβ1-40, total tau, NFL and GFAP levels were determined using a multiplex array (Neurology 4-Plex A Advantage Kit, Neurology 3-Plex A Advantage Kit. Quanterix, USA). pTau217 and pTau181 levels were measured using the following arrays (pTau217, ALZpath V2, P-tau181 V2.1, Quanterix, USA). The Aβ42/40 represented the ratio between Aβ1-42 and Aβ1-40.

Voxel-based morphometry

Structural T1-weighted MRI scans were processed using SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/). Images were segmented into gray matter, white matter, and CSF; spatially normalized to MNI space using Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra algorithm; modulated to preserve volume; and smoothed with an 8 mm FWHM Gaussian kernel. Voxel-wise comparisons between groups were performed using a general linear model with age, sex, and total intracranial volume as covariates. Statistical significance was defined as p < 0.05 with family-wise error correction for multiple comparisons and a minimum cluster extent of 100 voxels.

Statistical analysis

Based on amyloid status or biological diagnosis, we categorized the groups into 2 or 3 classes. Group comparisons of continuous variables were conducted using the Kruskal-Wallis test of the nonparametric method, followed by the Mann-Whitney test for the significant pairs comparison. Categorical variables were analyzed using the chi-square test. Pearson correlation coefficients were used to assess associations between continuous variables. The performance of individual biomarkers versus the combined panel was assessed using receiver operating characteristic (ROC) curves, and the areas under the curve (AUCs) were compared using the DeLong test. All statistical analyses were performed using R software (R Foundation for Statistical Computing, Vienna, Austria). A p-value < 0.05 was considered statistically significant. Additionally, we included 4 machine learning models to assess the performance of combined panel on group classifications.

Multiclass LR – Multiclass LR was employed in our machine learning model to handle 2 or 3 outcome categories. Compared to building multiple binary classifiers (e.g., One-vs-Rest strategy), multiclass LR offers a more unified and often more efficient solution. The multiclass LR model establishes regression relationships between independent variables and the probability of each category, directly quantifying how a single marker compares to a weighted panel. It substitutes the likelihood function of the data to estimate regression coefficients and then calculates prediction probabilities. The difference of Akaike information criterion (AIC) between the reduced model and the full model, △AIC, confirmed the feature importance.

Classification Tree – This model creates parent nodes by dividing categorical or continuous independent variables into distinct ranges and subtracting them into child nodes in subsequent stages. A tree is formed, dividing the sample space into different regions based on independent variables, which are then used for classification.

Random Forest – The Random Forest algorithm combines the predictions of multiple decision trees through averaging (for regression) or majority voting (for classification) to improve overall accuracy and reduce overfitting. Since its performance is influenced by the choice of hyperparameters, optimal values were determined using grid search with 5-fold cross-validation on the training set, allowing a balance between model complexity and variance. The key hyperparameters tuned in this study included the number of trees (n_estimators, range: 1–1000), maximum tree depth (max_depth, range: 1–100), and the minimum number of samples required at a leaf node (min_samples_leaf = 1).

SVM – SVM identifies a hyperplane in the sample space that maximizes the distance between two data types. It uses margin maximization and kernel transformations to capture potential non-linear biomarker interactions, ensuring that the findings of pTau217’s primacy holds even under non‐linear decision boundaries.

We selected these four classification algorithms that span linear, tree-based, ensemble, and kernel methods. This comprehensive approach allows us to test whether the dominance of pTau217 over a multi-marker panel is an artifact of any one algorithm or a robust phenomenon across diverse modeling frameworks. Moreover, the sensitivity, specificity, AUC, and precision are reported for all four models. Accuracy was first evaluated under different conditions: with and without including clinical data in the model.

To mitigate multicollinearity among our six plasma biomarkers (pTau217, pTau181, Aβ42/40, NFL, GFAP, total tau), we first standardized each marker to zero mean and unit variance and then performed principal component analysis (PCA). The first six principal components (PC1–PC6), explaining total variance, were used as orthogonal inputs for model development; comparisons with LASSO-regularized logistic regression on raw biomarkers produced confusion matrices within 1% overall accuracy, confirming robustness to collinearity.

Although both random split (60% training, 20% testing, and 20% validation) and 5-fold cross-validation methods were used to evaluate model performance, we emphasized the robustness of 5-fold cross-validation in our analysis as it ensures that every data point is used for both training and testing across different iterations. The results of a random split can vary substantially depending on how the data is partitioned, potentially leading to biased or unstable outcomes.

Combined-panel equations

We constructed each composite-panel score by applying the biomarker coefficients derived from our multivariable logistic regression models to the measured plasma concentrations. To ensure full transparency and reproducibility, we provide below the exact linear predictors used for each classification task; each score corresponds to the log‐odds index for amyloid status or clinical‐stage stratification:

1. A⁺ versus A⁻ (Stage 1–6)

   

2. A⁺ versus A⁻ (Stage 1–3)

   

3. Stage 1–3 versus 4–6 within A⁺

   

In each equation, the sign and magnitude of the coefficient indicate the direction and strength of that biomarker’s contribution to the log-odds of the specified outcome, where the coefficients were obtained from multivariable logistic regression models fitted to the training data for each classification task. Because these β-values vary across tasks (e.g., A⁺ vs. A⁻ across all stages, early-stage only, or stage progression), the relative influence of each marker changes. Overall, A higher positive score corresponds to a greater likelihood of amyloid positivity or advanced clinical stage.

In contrast, classification trees and random forests learn split rules on the raw features during training and assess feature importance post hoc (e.g., via mean decrease in Gini), rather than encoding it in a fixed formula. SVM similarly do not require a pre-specified weighting equation. A linear-kernel SVM yields a weight vector after fitting (the hyperplane coefficients), but this vector is a model output rather than an input. Kernel SVMs, meanwhile, have no direct weight vector in the original feature space and rely on support vectors and kernel transformations to define their decision boundaries.

Results

Demographics

After applying inclusion and exclusion criteria, 371 participants completed the study procedure (Table 1; Fig. 1). Post-hoc analysis revealed that the AD group had significantly higher levels of all plasma biomarkers than the CU group. Specifically, pTau217, pTau181, and GFAP levels were significantly higher, while Aβ42/40 was lower in AD patients compared to the non-AD group. Additionally, GFAP, NFL, and total tau levels were significantly higher in the non-AD continuum than in the CU group.

Table 1.

The descriptive statistics of patients for 3 biological categories

	AD (n = 143)	CU (n = 69)	Non-AD continuum (n = 159)	F, P values†
Male number, %	47, 32.9*#	37, 53.6	80, 50.3	12.34, 0.002
ApoE4 carriers, %	77, 53.9*#	12, 17.4	29, 18.2	52.13, < 0.001
MMSE	17.3, 11.0*#	27.0, 4.0	21.3, 7.0*	104.27, < 0.001
Amyloid CL	68.6, 44.3*#	-0.97, 8.9	-3.2, 16.2	208.04, < 0.001
Global Tau severity score	88.7, 30.8*#	-0.009,13.5	17.5, 25.73*	264.98, < 0.001
Age (year)	72.96, 9.5*	66.2, 12.0	71.4, 10.0*	26, < 0.001
Education	8.3, 6.0*	10.5, 8.0	9.2, 6.0	10, 0.007
pTau217 (pg/ml)	0.99, 0.55*#	0.27, 0.13	0.36, 0.16	211.04, < 0.001
Aβ42/40	0.036, 0.007*#	0.043, 0.009	0.042, 0.008	55.46, < 0.001
NFL (pg/ml)	28.08, 16.66	15.70, 10.41	33.90, 20.99*	45.35, < 0.001
GFAP (pg/ml)	561.9, 277.4*#	226.7, 109.7	325.3, 175.0*	134.27, < 0.001
pTau181(pg/ml)	40.07, 15.33*#	21.45, 10.24	24.13, 10.07	138.95, < 0.001
Total Tau(pg/ml)	5.84, 2.46*	4.38, 1.03	5.41, 2.10*	23.57, < 0.001

Abbreviations: APOE, apolipoprotein E; AD: Alzheimer’s disease; CU, cognitive unimpaired control; CL: centiloid; GFAP, glial fibrillary acidic protein; MMSE, Mini-Mental State Examination; NFL, neurofilament light chain

Diagnosis of non-AD continuum included amyloid negative (A), Tau (T) PET positive tauopathies (A-T+, n=60), progressive supranuclear palsy (n=28, A-T+) and neurodegenerative disorders with negative A and T status (A-T-, n=71)

Continuous data are expressed as mean, interquartile ranges† using the Kruskal-Wallis test. Post-hoc analysis using Bonferroni correction, with * indicates p<0.05 compared with control and # as p<0.05 compared with the non-AD continuum

Fig. 1.

(A) Voxel-based morphometry results among Alzheimer’s Disease (AD), cognitive unimpaired (CU) controls and other non-AD neurodegenerative disorders. The color bar represents T value. Significance set at p < 0.05 with family wise error corrections for multiple comparisons and cluster size 100. (B) The average image of [F18]Florzolotau PET images in 3 groups. The color bar represents standardized uptake value ratio

Among those with AD, 54.5% had MCI, 26.6% had mild dementia, and 18.9% had moderate dementia. In the non-AD group, the proportions of MCI, mild dementia, and moderate dementia were 71.1%, 22.0%, and 6.9%, respectively. Across all participants, MMSE scores and clinical stages followed a linear regression pattern (MMSE=-4.896 × clinical stage + 34.56, r = 0.83, R² = 0.69, p < 0.001); for each additional clinical stage, MMSE decreases by 4.896. Furthermore, MMSE scores and clinical stages in AD patients followed a linear trend (MMSE= -5.560 ×clinical stage + 36.62, r = 0.82, R² = 0.68, p < 0.001); for each additional clinical stage, MMSE decreases by 5.560.

In AD patients, MMSE scores correlated linearly with pTau217 (p = 0.001), GFAP (p = 0.0006), pTau181 (p = 0.02), NFL (p = 0.002), and total tau (p = 0.002, Supplementary Fig. 2), while disease duration (years) was positively correlated with NFL and GFAP levels (Supplementary Fig. 3). Consistent with plasma biomarker stratification, VBM analysis revealed significant gray-matter volume reductions in the hippocampus and posterior cingulate cortex in A + versus A– subjects (Fig. 1A). For non-AD group, the atrophic regions were localized in the anterior cingulate and hippocampal regions. The tau burden in AD were found in the temporal-parietal and the temporal regions (Fig. 1B).

Diagnostic accuracy using the LR model

Using a simple LR model, we compared the accuracy of single biomarkers versus a combined-panel approach across three clinical scenarios. The AUC values for the combined panel and individual biomarkers in distinguishing A + from A − status are shown in Fig. 2. Table 2 reports the cross-validated diagnostic performance of various logistic‐regression models—in three different clinical scenarios—using both raw biomarkers and their PCA‐derived composites. The full combined panel consistently achieves the highest AUC (0.946–0.949) and accuracy (0.889–0.911) in the A⁺ vs. A⁻ tasks, with only modest drops when using PCA‐derived panels (ΔAUC ≈ 0.01–0.02; Δaccuracy ≈ 0.02–0.03, Table 2, left panel; Fig. 2A). Accuracy of the combine panel further increased to 0.911 when the analysis was restricted to participants at stages 1–3 (Table 2, middle panel; Fig. 2B).

Fig. 2.

Receiver operating characteristic (ROC) analysis for evaluating the predictive performance of each biomarker for amyloid positivity response in a heterogenous groups. (A) Logistic regression model for binarize amyloid positive (A+) with amyloid negative (A-) in all cases (n = 371). (B) Individuals included cognitive unimpaired (CU) controls or cases with clinical staging of 1 to 3. GFAP: glial fibrillary acidic protein, NFL: neurofilament light chain. The combined panel are pTau217, pTau181, GFAP, NFL, total tau and Aβ42/40

Table 2.

The area under the ROC curve by binary logistic regression analysis (2 classes classification)

Clinical scenario		A + versus A- (stage 1–6)		A + versus A- (stage 1–3)		Stage 1–3 vs. 4–6 in A+
Cases number		n = 371		n = 258		n = 143
Variable	Model	AUC	accuracy	AUC	accuracy	AUC	accuracy
Combined-panel	a	0.949	0.895	0.949	0.911	0.608	0.594
Combined-panel (by 5 PCA)	a	0.938	0.871	0.948	0.891	0.596	0.573
pTau217	a	0.944	0.865	0.947	0.876	0.651	0.566
pTau181	a	0.859	0.784	0.855	0.771	0.580	0.573
GFAP	a	0.834	0.771	0.821	0.802	0.605	0.559
pTau217 & pTau181	a	0.941	0.860	0.931	0.891	0.637	0.566
pTau217 & pTau181 (PCA)	a	0.912	0.825	0.914	0.837	0.625	0.566
pTau217 & GFAP	a	0.939	0.863	0.944	0.872	0.642	0.601
pTau217 & GFAP (PCA)	a	0.914	0.860	0.923	0.872	0.644	0.601
pTau181 & GFAP	a	0.866	0.817	0.860	0.837	0.591	0.559
pTau181 & GFAP (PCA)	a	0.873	0.827	0.868	0.829	0.607	0.566
Combined-panel	b	0.946	0.889	0.945	0.911	0.567	0.566
Combined-panel (by 6 PCA)	b	0.932	0.871	0.935	0.888	0.575	0.545
pTau217	b	0.942	0.871	0.942	0.864	0.604	0.552
pTau181	b	0.855	0.782	0.839	0.779	0.536	0.566
GFAP	b	0.825	0.760	0.806	0.795	0.549	0.539
pTau217 & pTau181	b	0.939	0.879	0.927	0.876	0.594	0.573
pTau217 & pTau181 (PCA)	b	0.910	0.822	0.893	0.829	0.605	0.559
pTau217 & GFAP	b	0.937	0.871	0.939	0.868	0.605	0.580
pTau217 & GFAP (PCA)	b	0.897	0.852	0.895	0.872	0.631	0.587
pTau181 & GFAP	b	0.861	0.819	0.849	0.833	0.550	0.524
pTau181 & GFAP (PCA)	b	0.844	0.811	0.825	0.810	0.593	0.566

Accuracy is computed by the cut probability = 0.5

2 classes indicated amyloid positive (A+) or negative (A-) status

ROC: receiver operating characteristic, AUC: area under the ROC curve, PCA: Principal component analysis, GFAP, glial fibrillary acidic protein; NFL, neurofilament light chain

Model a includes only the listed test variable

Model b includes test variable and age, gender, educational years

A + versus A- (stage 1–6):

Combined-panel equation = 6.318pTau217) + 0.022pTau181)-54.52 (Aβ42/40)-0.072NFL) + 0.003GFAP) -0.267 (Total tau). Combined-panel (by PCA5 or PCA6) (scores computed on the first 5 or 6 principal components)

(I) By PCA_Combined-panel equation = 1.747(pTau217) + 1.610pTau181) − 0.652(Aβ42/40) – 3.071NFL) + 0.942GFAP) − 0.617(Total tau)

(II) By PCA_Combined-panel equation = − 0.875PC1–2.676PC2 + 2.364PC3 + 1.804PC4–0.039PC5

A + versus A- (stage 1–3):

Combined-panel equation = 12.071pTau217)-0.042pTau181)-57.85 (Aβ42/40)-0.065NFL) + 0.003GFAP) -0.038 (Total tau)

(I) By PCA Combined-panel equation = 1.95pTau217) − 1.459pTau181) − 0.579(Aβ42/40) – 3.903NFL) + 0.758GFAP) − 0.036(Total tau)

(II) By PCA Combined-panel equation = 0.573PC1–2.915PC2–2.283PC3 + 0.414PC4 + 2.806PC5

Stage 1–3 vs. 4–6 in A+:

Combined-panel equation = 1.51pTau217)-0.019pTau181) + 26.322 (Aβ42/40) + 0.006NFL) + 0.001GFAP) + 0.004 (Total tau)

(I) By PCA Combined-panel equation = 0.202pTau217) − 0.175pTau181) + 0.173(Aβ42/40) + 0.123NFL) + 0.371GFAP) + 0.0004(Total tau)

(II) By PCA Combined-panel equation = − 0.375PC1 + 0.24PC2 + 0.175PC3–0.045PC4 + 0.152PC5

pTau217 alone comes very close to the panel (AUC 0.942–0.947; accuracy 0.864–0.876), highlighting its strong individual predictive power. Two-marker combinations (especially pTau217 + GFAP or pTau217 + pTau181) slightly outperform pTau217 alone but remain below the full panel (ΔAUC ≈ 0.002–0.008). Their PCA versions again modestly lower performance.

All models perform less well on the stage-progression task (AUCs 0.536–0.651, accuracies 0.552–0.601, Table 2 right panel), reflecting the greater challenge of discriminating early versus late clinical stages. Including demographics (Model b) does not improve—and in fact slightly reduces—both AUC (ΔAUC ≈ − 0.003) and accuracy (Δaccuracy ≈ − 0.006), suggesting that age, gender, and education add little incremental signal beyond the biomarkers themselves.

Classification models and diagnostic accuracy across all stages

Using the multi-class LR and three predefined ML algorithms, we compared the diagnostic accuracy of the combined panel (Table 3). Two-class classification (A + vs. A−) consistently outperformed three-class classification (AD, non-AD, CU), yielding higher accuracy across all models. The multi-class LR and random forest models using the combined panel achieved accuracy above 90%, while the inclusion of demographic data did not improve accuracy. Contrary to expectation, the inclusion of demographic covariates (age, education, gender and MMSE) in the combined biomarker panel yielded a marginal decline in cross-validated AUC and accuracy. The best classification performance for the random forest model was achieved with the following hyperparameters: number of trees = 86, maximum depth = 67, and minimum samples per leaf = 1.

Table 3.

Classification methods for diagnostic accuracy for A + and A- in a heterogenous group

Methods	Class size	Model		Accuracy (%)
			5-fold cross-validation†		95% confidence interval	Random split‡
Multiclass Logistic Regression	2	Combined model	92.18 ± 2.00		(90.43, 93.93)	85.14
		Combined model (+a)	91.10 ± 2.64		(88.79, 93.41)	85.14
		pTau217	91.92 ± 1.86		(90.29, 93.55)	89.19
		pTau217 (+a)	90.57 ± 2.34		(88.52, 92.62)	83.78
	3	Combined model	73.05 ± 3.91		(69.62, 76.48)	71.62
		Combined model (+a)	74.38 ± 2.81		(71.92, 76.84)	70.27
		pTau217	68.99 ± 3.39		(66.02, 71.96)	63.51
		pTau217 (+a)	72.77 ± 1.16		(71.75, 73.79)	70.27
Random Forest	2	Combined model	92.46 ± 1.78		(90.90, 94.02)	86.49
		Combined model (+a)	91.91 ± 0.99		(91.04, 92.78)	89.19
		pTau217	88.95 ± 3.74		(85.67, 92.23)	82.43
		pTau217 (+a)	91.11 ± 1.50		(89.80, 92.42)	90.54
	3	Combined model	76.02 ± 3.10		(73.30, 78.74)	74.32
		Combined model (+a)	77.10 ± 2.03		(75.32, 78.88)	75.68
		pTau217	61.73 ± 1.14		(60.73, 62.73)	59.46
		pTau217 (+a)	74.67 ± 3.38		(71.71, 77.63)	70.27
Support Vector Machine	2	Combined model	88.94 ± 2.44		(86.80, 91.08)	86.49
		Combined model (+a)	87.59 ± 3.53		(84.50, 90.68)	85.14
		pTau217	92.46 ± 1.78		(90.90, 94.02)	89.19
		pTau217 (+a)	88.42 ± 2.75		(86.01, 90.83)	87.84
	3	Combined model	72.77 ± 1.78		(71.21, 74.33)	70.27
		Combined model (+a)	69.54 ± 1.18		(68.51, 70.57)	71.62
		pTau217	73.06 ± 2.36		(70.99, 75.13)	72.97
		pTau217 (+a)	72.23 ± 2.14		(70.35, 74.11)	70.27
Classification Tree	2	Combined model	89.50 ± 3.84		(86.13, 92.87)	86.49
		Combined model (+a)	90.30 ± 3.05		(87.63, 92.97)	87.84
		pTau217	91.92 ± 2.49		(89.74, 94.10)	89.19
		pTau217 (+a)	89.49 ± 3.08		(86.79, 92.19)	89.19
	3	Combined model	72.26 ± 3.91		(68.83, 75.69)	72.97
		Combined model (+a)	74.94 ± 2.91		(72.39, 77.49)	71.62
		pTau217	73.06 ± 2.16		(71.17, 74.95)	72.97
		pTau217 (+a)	74.94 ± 2.91		(72.39, 77.49)	72.97

GFAP, glial fibrillary acidic protein; NFL, neurofilament light chain

2 Classes: Amyloid positive (A+) versus negative (A-)

3 Classes: Alzheimer’s disease (AD), Cognitive unimpaired controls, non-AD continuum

Combined model: plasma biomarkers of pTau217, pTau181, Aβ 42/40, NFL, GFAP, Total tau

(+ a) with demographic factors of Age, Gender, Education, mini-mental status examination score

†: Average accuracy of 5-fold cross-validation

‡: Accuracy of randomly dividing the cases into training set (n = 223), validation set (n = 74) and testing set (n = 74)

For feature importance in the multi-class LR model, pTau217 exhibited the largest ΔAIC and coefficient for distinguishing A⁺ from A⁻ (Table 4). Age, gender, and education had relatively small coefficients and negative △AIC, suggesting they did not significantly improve model performance. The importance of pTau217 was consistently observed in the three-class model (AD, non-AD, CU, Table 4). In contrast, NFL played a more significant role in distinguishing non-AD from CU. In the random forest model, the mean decrease in Gini index also supported the importance of pTau217 in the combined model (Supplementary Table 1), both in two- and three-class classifications.

Table 4.

Importance of variables for diagnosis by multiclass logistic regression model (n = 371)

2 diagnostic classes			3 diagnostic classes
Variable	coefficient†	△AIC‡	Variable	AD coefficient†	Non-AD coefficient†	△AIC‡
pTau217	6.210 ± 1.247	41.247 ± 9.500	pTau217	6.118 ± 0.968	-0.071 ± 0.886	39.106 ± 8.331
NFL	-0.076 ± 0.011	25.223 ± 5.247	NFL	-0.029 ± 0.014	0.055 ± 0.013	32.979 ± 6.132
Total Tau	-0.326 ± 0.135	7.135 ± 6.258	MMSE	-0.393 ± 0.037	-0.354 ± 0.050	31.185 ± 5.012
Aβ42/40	-62.117 ± 16.420	5.329 ± 3.509	Total Tau	-0.211 ± 0.138	0.153 ± 0.129	7.257 ± 5.942
MMSE	-0.070 ± 0.015	2.871 ± 1.719	Aβ42/40	-68.578 ± 17.805	-4.306 ± 12.894	4.257 ± 3.412
GFAP	0.002 ± 0.001	1.346 ± 1.426	pTau181	-0.022 ± 0.030	-0.063 ± 0.012	2.288 ± 1.720
pTau181	0.020 ± 0.033	-0.036 ± 2.859	GFAP	0.005 ± 0.001	0.003 ± 0.001	1.821 ± 1.008
Gender	-0.487 ± 0.184	-0.604 ± 0.985	Education year	0.092 ± 0.046	0.087 ± 0.027	-1.060 ± 1.672
Age	0.018 ± 0.005	-1.595 ± 0.203	Gender	-0.808 ± 0.191	-0.404 ± 0.088	-1.781 ± 0.982
Education	0.008 ± 0.028	-1.747 ± 0.142	Age	0.037 ± 0.016	0.039 ± 0.014	-2.104 ± 1.320

3 Diagnostic classes: cognitive unimpaired (CU) controls, Alzheimer’s disease (AD), non-AD neurodegenerative disorder. The reference group is CU

2 diagnostic classes: A + versus A- group. The reference group is CU and non-AD

Abbreviations: AIC: Akaike information criterion; GFAP, glial fibrillary acidic protein; MMSE, Mini-Mental State Examination; NFL, neurofilament light chain;

AIC = 2k-log(L), where k is the number of parameters in the model, and L is the maximum value of the model likelihood function

†: Multiclass Logistic Regression model’s average coefficient and standard deviation for 5-fold cross-validation

^‡ : The average and standard deviation of AIC difference for 5-fold cross-validation of logistic regression training models

As pTau217 emerged as the most important feature, we compared its accuracy with that of the combined panel (Table 3). The classification accuracy of pTau217 was comparable to that of the combined model in two-class models, regardless of the inclusion of demographic data. pTau217 alone achieved an accuracy of > 90% in multi-class LR, SVM and classification tree even without demographic data. Interestingly, in SVM and classification tree models, pTau217 alone outperformed the combined panel in two-class classifications.

The accuracy of other biomarkers is detailed in Supplementary Table 2. The performance matrix from five-fold cross-validation for A+/A − classification is presented in Supplementary Table 3. Although three-class classification did not achieve as high accuracy as two-class classification, ML models demonstrated strong potential for differential diagnosis. The combined panel achieved 94.4% accuracy in distinguishing AD from CU and 91.7% accuracy in differentiating AD from the non-AD continuum (Supplementary Table 4). The results for pTau217 and other single biomarkers in AD vs. CU and AD vs. non-AD continuum classification are shown in Supplementary Tables 5 and 6, respectively.

Diagnostic accuracy at early clinical stages

We further evaluated the accuracy of the combined panel using different classification models for diagnosing AD in an earlier clinical-stage cohort (stages 1–3, n = 258, Table 5). In this setting, the multi-class LR model achieved the highest accuracy, with the combined panel reaching 94.23% accuracy in random split validation and 93.42 ± 1.72% in five-fold cross-validation. By comparison, pTau217 alone achieved an accuracy of 94.4%, sensitivity of 94.4%, specificity of 88.1%, and AUC of 94.8% in predicting amyloid-positivity (5-fold cross-validation), demonstrating that a single biomarker can nearly match the full panel’s performance (Supplementary Table 3). The performance of other plasma biomarkers is shown in Supplementary Table 7.

Table 5.

Classification methods for diagnostic accuracy for amyloid positivity at earlier clinical stages

Methods	Features	Accuracy (%)
		5-fold cross-validation†	95% confidence interval	Random split‡
Binary Logistic Regression	pTau217	93.02 ± 4.26	(89.29, 96.75)	90.38
	pTau217 (+ a)	91.46 ± 3.82	(88.11, 94.81)	90.38
	Combined panel	93.42 ± 1.72	(91.91, 94.93)	94.23
	Combined panel (+ a)	91.86 ± 4.20	(88.18, 95.54)	88.46
Random Forest	pTau217	87.98 ± 2.18	(86.07, 89.89)	84.62
	pTau217 (+ a)	93.41 ± 4.05	(89.86, 96.96)	80.77
	Combined panel	91.86 ± 4.20	(88.18, 95.54)	78.85
	Combined panel (+ a)	91.07 ± 5.28	(86.44, 95.70)	90.38
Support Vector Machine	pTau217	92.62 ± 4.25	(88.89, 96.35)	92.31
	pTau217 (+ a)	88.76 ± 2.84	(86.27, 91.25)	84.62
	Combined panel	87.16 ± 7.22	(80.83, 93.49)	80.77
	Combined panel (+ a)	89.52 ± 2.69	(87.16, 91.88)	86.54
Classification Tree	pTau217	89.54 ± 4.69	(85.43, 93.65)	90.38
	pTau217 (+ a)	89.93 ± 3.15	(87.17, 92.69)	90.38
	Combined panel	89.92 ± 5.08	(85.47, 94.37)	82.69
	Combined panel (+ a)	88.38 ± 6.58	(82.61, 94.15)	80.77

Combined panel: plasma biomarkers of pTau217, pTau181, Aβ 42/40, NFL, GFAP, Total tau

(+ a) with demographic factors of Age, Gender, Education, mini-mental status examination score

†: Average accuracy of 5-fold cross-validation

‡: Accuracy of randomly dividing the cases into training set (n = 154), validation set (n = 52) and testing set (n = 52), total cases 258

Cases enrolled were Alzheimer’s disease (AD), non-AD (at clinical stage 1–3), and Cognitive unimpaired controls

Abbreviations: GFAP, glial fibrillary acidic protein; NFL, neurofilament light chain

To quantify MMSE’s unique contribution, we compared multinomial LR models with and without MMSE. The model including MMSE yielded a ΔAIC of 4.2 compared to the model excluding MMSE, indicating only modest improvement in fit. This confirms that our staging predictions are not driven solely by MMSE scores.

To investigate the impact of clinical covariates on model performance, we analyzed the importance of features in the multi-class LR model. Among clinical variables, pTau217 had the highest ranking based on △AIC (Supplementary Table 8), followed by NFL, suggesting that pTau217 significantly contributed to classification accuracy in early-stage diagnosis.

Model performance in clinical stage segregation (stages 1–3 vs. 4–6) in AD

Finally, we restricted the analysis to individuals diagnosed with AD (n = 143) to evaluate whether the combined panel outperformed individual biomarkers in distinguishing early-stage (1–3) from later-stage (4–6) AD. Among all ML models, the multi-class LR model (Table 6) achieved the best performance, and the inclusion of demographic data significantly improved model accuracy. In five-fold cross-validation, model accuracy using either the isolated biomarker or the combined panel with demographic variables ranged from 87.36 to 89.53%. Results from the other 3 ML models are presented in Supplementary Table 9.

Table 6.

Classification performances in alzheimer’s disease for early or late staging prediction

Methods	Feature	Accuracy(%)
		5-fold cross-validation†	95% confidence interval	Random split‡
Binary Logistic Regression	pTau181	65.05 ± 3.22	(62.23, 67.87)	58.62
	pTau181 (+ a)	88.82 ± 4.52	(84.86, 92.78)	86.21
	Aβ42/40	53.87 ± 4.26	(50.14, 57.60)	37.93
	Aβ42/40 (+ a)	89.53 ± 4.18	(85.87, 93.19)	89.66
	NFL	67.14 ± 7.59	(60.49, 73.79)	48.28
	NFL (+ a)	87.39 ± 6.46	(81.73, 93.05)	79.31
	GFAP	68.55 ± 4.16	(64.90, 72.20)	55.17
	GFAP (+ a)	89.53 ± 4.18	(85.87, 93.19)	75.86
	pTau217	69.16 ± 7.16	(62.88, 75.44)	68.97
	pTau217 (+ a)	87.41 ± 3.09	(84.70, 90.12)	82.76
	Total Tau	61.45 ± 7.16	(55.17, 67.73)	37.93
	Total Tau(+ a)	88.82 ± 4.52	(84.86, 92.78)	82.76
	Combined panel	66.43 ± 3.10	(63.71, 69.15)	48.28
	Combined panel (+ a)	87.36 ± 5.49	(82.55, 92.17)	79.31

Combined panel: plasma biomarkers of pTau217, pTau181, Aβ 42/40, NFL, GFAP, Total tau

(+ a) with demographic factors of Age, Gender, Education, mini-mental status examination score

†: Average accuracy of 5-fold cross-validation

‡: Accuracy of randomly dividing the cases into training set (n = 85), validation set (n = 29) and testing set (n = 29), total cases 143

Cases enrolled were amyloid-positive Alzheimer’s disease, with binarized outcomes (clinical stage 1–3 versus 4–6)

Abbreviations: GFAP, glial fibrillary acidic protein; NFL, neurofilament light chain

To explore why the inclusion of demographic factors improved biomarker performance, we assessed the coefficient and the △AIC values (Supplementary Table 10). MMSE had the highest ranking (△AIC = 78.74) among all features, indicating its significant contribution to clinical stage prediction.

Discussions

This study explored a cost-effective search model for identifying patients with amyloid positivity using ML algorithms and plasma biomarkers. The findings highlight three major insights. First, findings from the 4 models demonstrated the high accuracy of combined model, with pTau217 identified as the most important feature. Second, when the clinical condition was at an earlier stage, both the combined panel and pTau217 alone achieved high accuracy. Finally, MMSE was the highest-ranked predictor among all features for clinical staging, consistent with its established role as a practical staging tool for clinicians. Notably, pTau217 achieved higher diagnostic accuracy in the earlier stages of disease, highlighting the complementary value of combining clinical staging with plasma biomarker assays. Together, our findings support the integration of clinical staging with plasma biomarkers to improve diagnostic precision in AD, providing independent validation in a Taiwanese memory-clinic cohort.

The emergence of plasma biomarkers in AD diagnosis represents a transformative advancement in the pursuit of accessible and scalable diagnostic strategies. Our study demonstrated that a combined plasma biomarker panel achieved the high diagnostic accuracy (> 92%) for detecting amyloid positivity in a heterogeneous population. Notably, the diagnostic performance further improved in early-stage individuals, with accuracy increasing to 93.4% among participants in clinical stages 1–3. These results highlight the importance of timing in utilizing the plasma panel to identify AD in real-world populations.

Our findings also indicate that pTau217 and the combined panel have comparable accuracy. Integrating plasma biomarkers, particularly pTau217, into clinical workflows for AD diagnosis is an active area of research. Prior studies have demonstrated its high sensitivity and specificity in distinguishing AD from non-AD conditions [27, 28], and our data shows that pTau217 alone achieves diagnostic accuracy exceeding 90%, both in heterogeneous and early-stage populations. Furthermore, pTau217 had the highest rank both in the SVM or classification tree model, further confirming its superior predictive value over other plasma biomarkers. While the combined-panel approach yields marginal improvements in classification metrics, its greater cost or operational complexity may limit adoption in routine clinical workflows. Meanwhile, the inclusion of demographic variables to the combined panel results in decrease performance. This likely reflects multicollinearity between demographic factors and biomarker levels, as well as sample-size constraints, indicating that our biomarker measurements alone capture the majority of predictive signal in this cohort. While the role of fluid biomarkers in routine clinical practice is still evolving, amyloid PET imaging remains the gold standard for confirming amyloid positivity—a key criterion for eligibility in disease-modifying treatments such as lecanemab and donanemab [29, 30].

We found that the multiclass LR model may outperform SVMs and classification trees and perform equivalently to random forests in diagnosing AD, especially when applied to a combined biomarker panel. We also recognize that the relative performance of SVM, and tree-based approaches depends in part on their hyperparameter settings—such as regularization parameters in SVM and maximum depth or number of trees in ensemble methods—and that systematic tuning or default choices may influence comparative accuracy [31, 32]. While LR is a generalized linear model that estimates the probability of class membership by modeling the relationship between independent variables and the log-odds of the dependent variable, this approach is particularly effective when the relationship between predictors and the outcome is approximately linear. As biomarkers such as pTau217, NFL, and MMSE already have strong predictive power, LR efficiently assigns weights without needing complex non-linear modeling. We also handled the multicollinearity through regularization to ensure stable coefficients even when biomarkers are highly correlated. The use of AIC helps optimize feature selection in LR. This transparency is particularly valuable in clinical settings, facilitating informed decision-making by healthcare professionals.

Cognitive testing remains a cornerstone in the clinical staging of AD, offering direct insight into functional impairment and disease progression. This study used clinical staging based on functional evaluation to determine the optimal timing for plasma biomarker assessment. Our study revealed significant correlations between NFL or GFAP with MMSE scores and were associated with disease duration in AD. However, these markers do not capture the functional domains assessed by clinical evaluations. While biomarkers such as NFL and GFAP reflect neurodegeneration and astroglia activation [33], they may align more closely with biological staging frameworks and are not direct substitutes for clinical evaluation. Thus, while biological and clinical staging are highly related, complex interactions with factors such as age, gender and brain reserve still exist [20, 34].

Our study had several limitations. First, this study was designed using cross-sectional data. Future studies incorporating longitudinal follow-ups will validate whether pTau217 remains stable in AD pathological trajectory and how its diagnostic accuracy evolves throughout disease progression. However, the major strength of our study lies in the temporal design, where blood samples were collected prior to the amyloid PET imaging, allowing for a realistic simulation of a pre-screening strategy. This enhances the cost-effectiveness of plasma biomarker-based models by reflecting how these tools might be used in practice to triage patients for confirmatory imaging or therapeutic eligibility. Second, while our models demonstrated high accuracy, the sample size is still small and their generalizability to diverse clinical settings remains uncertain. Participants in this study were primarily drawn from the memory clinics, where the prevalence of amyloid pathology is higher than in general population screenings. The performance of pTau217-based screening in primary care settings or asymptomatic individuals needs further investigation. Third, plasma assays for pTau217 and other markers have not yet received full FDA approval, and inter-assay variability or platform-specific differences still require more evidence.

In conclusion, our data show that pTau217 alone achieves > 92% accuracy while reducing assay costs. We further demonstrate that measuring biomarkers within six months of PET strikes the optimal balance between accuracy and expense. These findings establish an empirically grounded framework for optimizing both sampling timing and economic efficiency in plasma-based AD screening. Future health-economic modeling in community cohorts is warranted to validate clinical utility and implementation pathways.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

H.I and M.C wrote the initial draft, prepare the tables and figures. K.L, C.G, S.H, C.W, K.J performed data curation, data interpretation and analysis. C.C had critical revision and funding supports. All authors reviewed the manuscript.

Funding

This research was supported by the National Science and Technology Council, Grant/Award Numbers: NSTC113-2321-B-182 A-005, to CCC.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

The study was approved by the institutional review board of the Chang Gung Memorial Hospital and is conducted in accordance with the latest Declaration of Helsinki Manuscripts, including written informed consent from all participants. We confirmed that all human subjects provided informed consent for this study.

Competing interests

The authors declare no competing interests.