Introduction

Alzheimer’s disease (AD) is the most common type of dementia in the elderly, which is characterized by amyloid plaque comprised of amyloid-β (Aβ) and neurofibrillary tangles comprised of hyperphosphorylated tau [1,2,3]. Currently, disease-modifying therapies for AD are still lacking [4], and clinical trials of drugs targeting the pathological aspects have suffered serious setbacks, partially due to late intervention time and inaccurate clinical diagnosis [5]. Our group and others previously reported that 9–35% of patients clinically diagnosed with probable AD were Aβ negative [6,7,8,9], whereas 26–33% of cognitively normal elderly were Aβ positive in the brain [Clinical assessments and diagnosis of AD dementia

The clinical assessments and diagnosis of AD dementia were performed following our previous protocol [20]. In short, the demographic characteristics (including age, sex, education level), history of present illness, medical history (including diabetes, hypertension, dyslipidemia, coronary heart disease, etc.) and medication use were collected. Then, all participants underwent clinical assessments including physical examination, laboratory tests, APOE genoty**, magnetic resonance imaging, and neuropsychological tests. Diagnosis of AD dementia was made according to the criteria of the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorder Association (NINCDS-ADRDA) [21].

CSF sampling and processing

CSF samples were collected by lumbar puncture and processed according to a standard procedure [22]. Specifically, the CSF samples without visible blood contamination were collected in polypropylene tubes, followed by centrifugation at 2000 × g for 10 min at room temperature within 2 h. The supernatant was aliquoted and stored frozen at −80 °C until analysis.

Measurements of CSF biomarkers

CSF Αβ42 levels were determined using sandwich ELISA (INNOTEST® β-AMYLOID (1–42), Fujirebio, Belgium). CSF levels of total tau and P-tau181 were determined using sandwich ELISA INNOTEST hTau Ag, and INNOTEST PHOSPHO-TAU (181), respectively. All measurements were performed by an experienced laboratory technician who was blinded to the clinical information.

Statistical analysis

Sample size calculations were performed in PASS 11.0 software (NCSS, Kaysville, USA). In accordance with the estimation method of sample content for diagnostic test evaluation, we defined that Power (1-beta) = 0.95, alpha = 0.05, R (ratio of control to case group sample cases) = 2:1, AUC0(AUC to be achieved) = 0.7, AUC1(AUC from previous information) = 0.85, Type of data = continuous, Alternative Hypothesis = two-sided test. The results showed 52 cases in the case group and 104 cases in the control group.

The data are expressed as mean ± standard deviation (SD) or median (interquartile range, IQR) for numerical variables or as the count (%) for categorical variables. The differences in demographic characteristics and CSF biomarker levels between AD and control groups were assessed with two-tailed independent t-test, Mann Whitney U test, or χ2 test as appropriate. Spearman correlation analyses were used to examine the correlations between mini-mental state examination (MMSE) scores and CSF biomarkers levels.

The receiver operating characteristic curve (ROC) analysis is used to evaluate the diagnostic value of CSF biomarkers. The area under the curve (AUC), Akaike information criterion (AIC), sensitivity, specificity, accuracy, and the diagnostic cutoffs were estimated according to the largest Youden index. The combined diagnostic model of CSF biomarkers (we named it CSF index) was established by logistic regression (enter method), with Aβ42 (A), P-tau181 (T), and T-tau (N) as independent variables. Moreover, demographic information, including age (A’), sex (S), and APOE ε4 status (E), is added incrementally to the optimal model by logistic regression (forward method). Specifically, one demographic indicator is added at each step on the principle of the lowest AIC, and the process is repeated at the next step until the AIC does not decrease any further. AUC, AIC, and accuracy were calculated for each model. The DeLong test was used to compare the statistical significance between ROC curves [23]. Internal validation was performed by 1000 bootstrapped trials to evaluate the fitted degree among our apparent model, the Bias-corrected model, and the ideal model; the mean absolute error (MAE) < 0.05 meant high fitted degree.

All hypothesis testing was two-sided, p < 0.05 was defined as statistically significant. The computations were performed using SPSS 26.0 software (IBM SPSS Inc., Chicago, USA) and the R programming language (version 4.1.1).

Results

Characteristics of the study population

A total of 64 AD dementia patients and 105 age- and sex-matched cognitively normal (CN) controls were included in this study. The characteristics of these participants are shown in Table 1. There were no significant differences in age and sex between the two groups. The proportion of APOE ε4 carriers was higher in the AD group (p = 0.001). AD group had lower education levels and MMSE scores (p < 0.001).

Table 1 Characteristics of the participants.
Full size table

Cutoffs of core CSF biomarkers

Compared with controls, AD dementia patients had significantly lower Aβ42 levels, higher P-tau181 and T-tau levels in CSF (p < 0.001) (Fig. 1A, Table 1). The differences remained significant after adjusting for APOE ε4 status, education level, and comorbidities (p < 0.05). MMSE scores were positively associated with CSF Aβ42 (r = 0.665, p < 0.001), and negatively with P-tau181 (r = −0.451, p < 0.001) and T-tau (r = −0.557, p < 0.001) (Fig. 1B).

Fig. 1: Comparison of single CSF biomarkers and their correlation with MMSE scores.
figure 1

A Comparison of CSF Aβ42, P-tau181, and T-tau between AD (n = 64) and CN (n = 105) group. Boxes represent the 25th, 50th, and 75th data percentiles. Whiskers represent the lowest and highest data. The dashed lines indicate the cutoff values for each biomarker. B Correlations between CSF Aβ42, P-tau181, T-tau, and MMSE scores. The best-fit linear regression line is shown and 95% confidence intervals are superimposed. MMSE, mini-mental state examination; AD, Alzheimer’s disease; CN, cognitively normal.

Full size image

ROC analyses were performed to determine the diagnostic values of single CSF biomarkers. The cutoff value of CSF Aβ42 to distinguish AD from CN was 933 pg/mL (A+: Aβ42 < 933 pg/mL), with AUC of 94.0% (95%CI: 90.4–97.5%), the sensitivity of 89.1% and specificity of 87.6%. The diagnostic accuracy of Aβ42 was 88.2%. The cutoff values of P-tau181 and T-tau were 48.7 pg/mL and 313 pg/mL (T+: P-tau181 > 48.7 pg/mL; N+: T-tau>313 pg/mL), with AUC of 70.3% and 83.2%, respectively. The diagnostic accuracies were 69.2% for P-tau181 and 81.1% for T-tau, lower than that of Aβ42 (Table 2).

Table 2 Performance of CSF biomarker cutoffs.
Full size table

To further verify whether the cutoffs of CSF biomarkers can distinguish AD dementia patients from CN intuitively, we further analyzed their frequency distribution. As shown in Fig. 2A, the distribution of Aβ42 levels was in good agreement with the classification of the disease status, showing a bimodal distribution.

Fig. 2: Frequency distribution and internal validation of single CSF biomarkers.
figure 2

A Frequency distribution of CSF Aβ42, P-tau181, and T-tau. The dashed vertical lines indicate the cutoff value for each biomarker. B The bootstrap-validated of CSF Aβ42, P-tau181, and T-tau. The Y-axis indicates the actual probability of AD and the X-axis indicates the predicted probability of AD. The 45-degree black dotted line represents the ideal prediction; the solid black line surrounding the 45-degree black dotted line represents the bias-corrected prediction; the black dotted line surrounding the 45-degree black dotted line represents the apparent prediction. AD, Alzheimer’s disease; CN, cognitively normal; MAE, mean absolute error.

Full size image

Internal validation was performed using bootstrap** with 1000 repetitions to evaluate the reliability of CSF biomarkers. The results showed that Aβ42 and T-tau had a high fitted degree among our apparent model, the Bias-corrected model, and the ideal model (Aβ42: MAE = 0.013; T-tau: MAE = 0.024), while P-tau181 had a medium fit (MAE = 0.054) (Fig. 2B), indicating the reliability of the diagnostic efficacy of Aβ42 and T-tau.

Combined models of CSF biomarkers

To improve the accuracy of AD diagnosis, we established combined diagnostic models of CSF biomarkers, including AT, AN, TN, and ATN, by logistic regression (see Table 2 for details). Compared with the controls, AD group had significantly higher AT, AN, ATN indices, and lower TN index (p < 0.001) (Fig. 3A), even after adjusting for APOE ε4 status, education level, and comorbidities (p < 0.05). The frequency distributions of AT, AN, and ATN showed a good agreement with the classification of the disease status (Fig. 3B).

Fig. 3: Combined models of CSF biomarkers.
figure 3

A Comparisons of AT, AN, TN, and ATN model indices between AD (n = 64) and CN (n = 105) group. Boxes represent the 25th, 50th, and 75th data percentiles. Whiskers represent the lowest and highest data. The dashed lines indicate the cutoff value for each index. B Frequency distribution of AT, AN, TN, and ATN model indices. The dashed vertical lines indicate the cutoff values for each index. C ROC curves of CSF biomarkers and combined model indices. D AUC (x-axis) and AIC values (numbers in plots) for each biomarker and combined biomarker model. The dashed vertical line shows AUC = 0.7. E The OR values represent the contribution of each biomarker to the combined models. The error bars represent 95% confidence intervals. F The bootstrap-validated of CSF AN index. The Y-axis indicates the actual probability of AD and the X-axis indicates the predicted probability of AD. A, Aβ42; T, P-tau181; N, T-tau; AD, Alzheimer’s disease; CN, cognitively normal; ROC, receiver operating characteristic curve; AUC, area under the curve; AIC, Akaike information criterion; OR, odds ratio; MAE, mean absolute error.

Full size image

ROC analyses were performed to determine the AD diagnostic accuracy of each model; the lowest AIC, the best tradeoff between model fit and model complexity, was used to select the optimal model. As shown in Table 2 and Fig. 3, among the single biomarkers and combined biomarker models, CSF Aβ42 alone and the AN and the ATN models had higher and similar AUCs by DeLong test (p > 0.05); whereas the AN model had the lowest AIC, indicating the best diagnostic performance. The cutoff value of AN index to distinguish AD from control was −0.368, with an AUC of 94.9% (95%CI: 91.9–98.0%), a sensitivity of 90.6% and a specificity of 89.5%. The internal validation indicated that the AN model was reliable for AD diagnosis (MAE = 0.013) (Fig. 3F).

Integrative models of demographic characteristics with CSF biomarkers

Age, sex, and APOE genotype are associated with the risk of AD, so we investigated whether integrating demographic information could improve the diagnostic efficacy of CSF biomarker models. A data-driven model selection was performed to select the optimal model with the lowest AIC. The AN model, the best-combined biomarker model, was used as the basis; then age, sex, and APOE ε4 status were added in a stepwise procedure to examine the performance of integrative models. Better model performance was defined as being at least two AIC points lower than the previous model (ΔAIC > 2) [24] (Fig. 4A). The addition of demographic information slightly increased the AUCs and accuracy although no significant differences were detected by the DeLong test (p > 0.05) (Fig. 4B). The first step generated the ANA’ model (A’: age) with the AIC of 95.47 and AUC of 95.5%; The second step generated the ANA’E (E: APOE ε4 status) model, with AIC of 93.31 and AUC of 96.0% (95%CI: 93.2–98.9%). In the third step added sex, the AIC no longer decreased in ANA’ES (S: sex) model, with the higher AIC of 94.67 and AUC of 96.2%. Therefore, ANA’E had the lowest AIC in the above models, indicating the best diagnostic performance. The cutoff value of ANA’E model index was −0.401, able to well distinguish the two populations. The diagnostic accuracy of ANA’E model was up to 90.5% (Fig. 4D, E). The internal validation also confirmed the reliability of the ANA’E model (MAE = 0.019) (Fig. 4F).

Fig. 4: Integrative models of demographic characteristics with CSF biomarkers.
figure 4

A Model selection process. The data-driven model was selected with the lowest AIC (ΔAIC > 2). Based on the AN model, age, sex, and APOE ε4 status were added in a stepwise procedure, the model with lower AIC was obtained by adding one indicator at each step. B ROC Curves of integrated model indices. C OR values represent the contribution of each indicator to the integrated models. The error bars represent 95% confidence intervals. D Comparison of ANA’E index between AD (n = 64) and CN (n = 105) group. Boxes represent the 25th, 50th, and 75th data percentiles. Whiskers represent the lowest and highest data. The dashed lines indicate the cutoff value for ANA’E index. E Frequency distribution of ANA’E index. The dashed vertical lines indicate the cutoff values for ANA’E index. F The bootstrap-validated of CSF ANA’E index. The Y-axis indicates the actual probability of AD and the X-axis indicates the predicted probability of AD. A, Aβ42; T, P-tau181; N, T-tau; A’, age; E, APOE ε4 status; S, sex; APOE ε4, apolipoprotein E ε4 allele; ROC, receiver operating characteristic curve; AUC, area under the curve; AIC, Akaike information criterion; OR, odds ratio; AD, Alzheimer’s disease; CN, cognitively normal; MAE, mean absolute error.

Full size image

Discussion

In this study, we defined the cutoff values of CSF Aβ42, P-tau181, and T-tau for AD diagnosis (A+: Aβ42 < 933 pg/mL; T+: P-tau181 > 48.7 pg/mL; N+: T-tau > 313 pg/mL) in a Chinese cohort. Among these single biomarkers, CSF Aβ42 had the highest diagnostic accuracy of 88.2% in distinguishing AD patients from cognitively normal participants. Among the combined models of CSF biomarkers, AN was the simplest model while showing good diagnostic performance, with an accuracy of 89.9% (cutoff value > −0.368). In addition, it makes sense to integrate age and APOE ε4 status in the model to increase the accuracy (90.5%) and performance of the diagnosis.

The diagnosis of AD has now moved into the pathological phase with the inclusion of CSF biomarkers and amyloid PET in international guidelines [25, 26]. Although amyloid PET is the intuitive marker of amyloid pathology, it reflects the accumulation of sufficient amyloid to form an amyloid PET signal over many years. Whereas CSF biomarkers show the state of production and clearance of Aβ42 [27], and are likely to be positive early in the course of the disease before sufficient amyloid has accumulated, making it important in early diagnosis of AD [58]. The “X” represents biomarkers associated with synaptic damage, apoptosis, oxidative stress, neuroinflammation, neuroimmunity, mitochondrial dysfunction, and unrealized pathologies of AD [59]. An integrated model based on the ATXN framework could be applied not only for diagnosis, differential diagnosis, and prognosis, but also for the treatment and related trials of AD. Since the network of pathophysiology is complex and full of interconnections, all the dimensions in the framework should be involved in cocktail therapy. However, there are some challenges before widespread use of the ATXN framework. Large multicenter studies are still required to validate and standardize these biomarkers and their cutoffs, and the accuracy of biomarkers in the ATXN framework needs to be improved based on ultrasensitive technologies. Clarification of the interacting mechanisms of these biomarkers furthermore can provide the theoretical foundation for the application of the ATXN framework.

There are some limitations to this study. First, due to the difficulty of collecting CSF from AD dementia patients, the sample size of this study was relatively small. Even though internal validation has been performed, there’s still a need to expand the sample size in the external validations. Second, the participants enrolled were clinically diagnosed and lacked pathological evidence of Aβ-PET. Adequate validation in sufficient Aβ+ AD patients and Aβ− controls is highly needed before the findings of this study can move toward clinical practice, which requires further expansion of the sample size and inclusion of more stringent diagnostic criteria based on Aβ-PET in the follow-up. Finally, the assessment and external validation of the differential diagnostic ability is equally important before entering clinical practice, and we need to include more patients with other types of dementia to validate the differential diagnostic efficacy of the model in the future.

In conclusion, this study established the cutoff values of CSF biomarkers for AD diagnosis in a Chinese cohort, which is essential for the clinical application of AD biomarkers in Chinese population.