Methodological approaches to account for assay changes in longitudinal biomarker analysis: insights from Alzheimer’s blood biomarkers in the MEMENTO cohort

Abstract

Background

Longitudinal studies allow the modelling of disease progression through repeated measurement of health outcomes, such as biomarkers. Changes in measurement tools over time, due to logistical or financial constraints, may challenge the statistical modeling of outcome trajectories. This study aims to compare two methods for managing changes in blood biomarkers assays over time, in the context of modeling their longitudinal trajectories.

Methods

We analyzed data from 2299 individuals in the French MEMENTO cohort, focusing on two Alzheimer's disease blood biomarkers: 181-phosphorylated tau (p-tau181) and neurofilament light chain (NfL). Baseline blood samples were quantified using an initial assay kit in 2021, while samples collected at 2- and 4-year follow-ups with updated kits in 2023. Two approaches were applied to derive conversion equations for aligning measurements from the initial to the updated assay: (i) a bridging study, requiring biomarker quantification using both the initial and the updated assay in a subsample of individuals and (ii) Latent Process Models (LPM), which established links between the two assays as measures of the same latent process over age, using biomarker measurements available at the 3 timepoints. Prediction error rates were computed, and biomarker trajectories estimated with linear mixed models according to two variables of interest (education level, cognitive impairment).

Results

Prediction error rates were slightly higher for LPM than for bridging for both NfL and p-tau181. While the two methods yielded similar predictions around the median, discrepancies were observed at the tails of the distribution of the observed values. Longitudinal trajectories showed consistent associations for the variables of interest at baseline and during follow-up for both biomarkers.

Conclusions

LPM provide a feasible and efficient method for managing changes in biomarker quantification assays in longitudinal studies. LPM yields results comparable to traditional bridging studies without requiring additional sample analysis. This approach is particularly advantageous in studies with long-term follow-up, where changes in measurement tools cannot always be avoided, offering a straightforward and resource-efficient solution.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-026-02042-3.

Background

Longitudinal epidemiologic studies enable the modeling of disease trajectories and risk factors, through repeated biomarker measurements across multiple domains over time [1]. Measurement tools must provide consistent and reproducible results over the course of the study, to ensure that observed changes reflect the clinical condition under study rather than variations in measurement techniques [2]. In longitudinal studies, biological samples such as blood, cerebrospinal fluid, urine, saliva or microbiota are collected over time and stored in a biobank for subsequent biomarker assessment, either for discovery purposes or when budget allows. Ideally, repeated measurements of a given biomarker should be conducted under standardized conditions, using the same devices, assays, and operators within a short timeframe. However, achieving these conditions is challenging in long-term and/or large-scale studies, due to inherent constraints like human and financial resources, assay availability, and logistical factors.

In Alzheimer’s disease (AD), key pathological biomarkers are assessed through lumbar puncture or positron emission tomography (PET) imaging within specialized clinical settings [3]. The recent development of ultra-sensitive assays has enabled the measurement of these biomarkers in blood, making the exploration of their longitudinal dynamics a significant research focus due to their ease and rapid accessibility [4–6]. While these biomarkers remain the subject of extensive research to demonstrate their validity, some argue that they should be sufficient for AD diagnosis, whereas others suggest their use in well-defined clinical conditions in patients with cognitive impairment [7–9].

Changes in measurement techniques, analyzers, assays, kits or lots are common in biochemistry laboratories. When a new technique is a candidate to replace the current one, it requires a bridging study [10, 11]. A lot-to-lot validation (LTLV) study is needed before integrating a new lot [12–14]. Although the modalities of such studies may vary among clinical laboratories [15–17], they typically involve analyzing a series of samples with both measurement tools to derive a conversion equation, requiring significant human resources, biological samples, consumables, and time.

This issue parallels the challenge of changing assessment tools (e.g. questionnaires) across follow-ups, another obstacle that cohorts may face. As for biomarkers, the trajectory of the measure underlying the questionnaires provide insights on the disease progression. To address changes in assessment tool, statistical approaches such as latent variable models (LVM) or latent process models (LPM) have been proposed to estimate trajectories of cognitive performance [18], dietary habits [19], and physical activity [20]. A major advantage of these statistical approaches is that they do not require collecting additional data or using subsamples from biobanks. While LPM are suitable for addressing changes in biomarker measurement, it remains unclear whether they can serve as an alternative to bridging studies.

This study aims to compare two methods for accounting for assay changes over time when modeling longitudinal trajectories of two AD blood biomarkers. Using data from the French national cohort MEMENTO, we evaluated and compared the prediction errors of the two approaches, as well as their estimated biomarker trajectories over time.

Methods

Study population

The MEMENTO cohort [21] (NCT01926249) consecutively enrolled 2323 non-demented participants from 26 French memory clinic sites between 2011 and 2014. Participants were followed annually for up to 10 years. Alongside comprehensive clinical and neuropsychological evaluations, blood samples were collected at baseline (M0), month 24 (M24) and month 48 (M48), and stored in a centralized biobank at − 80 °C. A detailed description of the MEMENTO protocol and the enrolled sample characteristics has been published elsewhere [21]. The study protocol was approved by the Ethics Committee “CPP Sud-Ouest et Outre-Mer III”.

Assays used for biomarkers quantification

The 181-phosphorylated tau (p-tau181) peptide and the neurofilament light chain (NfL) were measured at M0, M24 and M48. These biomarkers are known to reflect AD neuropathological changes and neurodegeneration, respectively [4–6]. Elevated blood NfL concentrations indicate non-specific neuronal death, which is observed in neurodegenerative diseases such as AD [22–24], while increased blood p-tau181 concentrations are positively correlated with the presence of amyloid plaques in the brain, a hallmark of AD diagnosis [25–27].

Blood NfL and p-tau181 concentrations from M0, M24 and M48 samples were determined at the Bordeaux University Hospital research platform (PARS-Immunology team), using the same Quanterix HD-X analyzer throughout the study, following the manufacturer’s protocols. The assays used were commercial kits manufactured by Quanterix® and labeled for Research Used Only (RUO). Due to limited funding, M0 blood samples were analyzed in a single batch between May and September 2021, while M24 and M48 samples were analyzed between March and May 2023. Between these periods, the manufacturer introduced a change in the kits required to measure the two biomarkers (Table 1, Fig. 1).

Table 1.

Blood biomarker quantification. The MEMENTO cohort

	NfL	p-tau181
Whole cohort
M0
Period	May to September 2021	May to September 2021
Assay (manufacturer item reference)	NF-light Advantage Kit (103186)	p181-tau Advantage V2 Kit (103714)
Lot reference	502.607	502.613
N	2275	2077
Median (q1,q3), pg/mL	18.2 (13.4, 25.0)	0.9 (0.6, 1.4)
M24
Period	March to April 2023	March to May 2023
Assay (manufacturer item reference)	Neuro4Plex E (103670)	p181-tau Advantage V2.1 Kit (104111)
Lot reference	503.504	503.537
N	1820	1755
Median (q1,q3), pg/mL	22.9 (16.6, 31.8)	19.8 (14.7, 27.3)
M48
Period	March to April 2023	March to May 2023
Assay (manufacturer item reference)	Neuro4Plex E (103670)	p181-tau Advantage V2.1 Kit (104111)
Lot reference	503.504	503.537
N	1559	1533
Median (q1,q3), pg/mL	23.4 (17.0, 32.9)	20.6 (15.4, 27.9)
Bridging study
Period	October 2023	October 2023
Assay (manufacturer item reference)	Neuro4Plex E (103670)	p181-tau Advantage V2.1 Kit (104111)
Lot reference	503.819	503.769
N	56	32
Median (q1,q3), pg/mL	20.9 (15.6, 32.1)	26.1 (14.6, 52.7)
Lot-to-lot validation study
Period	October 2023	October 2023
Assay (manufacturer item reference)	Neuro4Plex E (103670)	p181-tau Advantage V2.1 Kit (104111)
Lot reference	503.819	503.769
N (duplicates)	79	52
Median (q1,q3), pg/mL	21.9 (16.7, 32.9)	38.0 (18.5, 75.1)

Fig. 1.

Overview of the methodology. A. Chronological summary of the analytical steps. B. Samples used according to the objective addressed. LOOCV: leave-one-out cross validation

Bridging study

To assess the impact of the kit change between 2021 (referred as the “initial” assay) and 2023 (referred as the “updated” assay), a subsample of M0 samples was analyzed with the updated assay in October 2023 (Fig. 1). For each biomarker, M0 concentrations from the initial assay were ranked in deciles, and a random subsample was drawn from each decile to ensure uniform coverage of the observed measurement range. The sample size was determined using the table provided by K. Linnet [10], based on a weighted Deming regression for method comparison, a proportional standard deviation, a type I error of 5% and 90% statistical power. The goal of the bridging analysis was to derive a conversion formula between the two assays. For the sake of comparison with future study results, we decided to predict the updated assay concentrations from the concentrations observed with the initial assay.

Although the assay versions remained the same, the kit lots changed between March 2023 (used for M24/M48 samples) and October 2023 (used for the bridging study) (Table 1, Fig. 1). To ensure consistency across lots, a lot-to-lot validation (LTLV) study was additionally conducted. The required sample size was based on Koh et al. [13], assuming a weighted Deming regression, a 1:1000 measurement range ratio, a 10% coefficient of variation and duplicate measurements A minimum of 40 samples was required to compare lots with a type I error of 5%. The mean of duplicate quantifications was used. If inconsistencies were detected across lots, an additional LTLV conversion equation was applied alongside the bridging equation.

Statistical analysis

Conversion equations using bridging

The conversion equation for the bridging study was derived using a weighted Deming regression (wDR). Unlike ordinary least squares regression (OLS), Deming regression (DR) accounts for measurement errors in the predictor variable in addition to the outcome variable, which is relevant in our dataset. We also evaluated four alternative regression models: unweighted DR, OLS, weighted least squares (wLS) and Passing-Bablock (PBR) [28]. PBR is a non-parametric approach, suitable for non-Gaussian distributions and robust to extreme values. The bridging equation was derived from the following model (1), estimated using a subset of individuals at the M0 visit:

1

Y_Updated,i and Y_Initial,i were the biomarker measures obtained with respectively the updated and the initial assay for an individual i at visit M0. The bridging equation was derived from the estimates and as .

For the LTLV study of the updated technique, the five regression models (OLS, wLS, DR, wDR and PBR) were tested, and the best model was selected based on the minimization of the leave-one-out cross validation (LOOCV) prediction error to optimize the detection of a difference between lots. LTLV equation was defined from the model (2) estimated on a subset of individuals at visits M24 and M48:

2

With and the biomarker measures obtained with respectively the new and old lot of the updated assay for an individual i at visit j = (M24, M48).

If we observed a significant difference in the two lots expressed in (2) by an estimated intercept different from 0 and/or an estimated slope different from 1, the LTLV conversion equation was subsequently applied on the bridging equation giving , where hat denotes the estimate.

The prediction errors of the bridging and LTLV candidate models were estimated using LOOCV due to the limited sample size.

Conversion equations using LPM

The Latent Process modeling approach relies on the core assumption that the different assays all measure the same unobserved quantity, in our case the true level of the biomarker. Unlike LVM, LPM accounts for repeated measures by modeling the underlying biomarker trajectory over time [29]. Given that the assays changed sequentially during follow-up, we used individuals’ age as the time scale to map the two assays. The latent process trajectory was modeled using a structural linear mixed model (LMM), with random effects on the intercept, age and age² to account for the correlation of the measures from a same individual. The latent process (Λ) was linked to the measurements obtained from both the initial (Y_Initial) and the updated (Y_Updated) assays using assay-specific linear link functions (3 and 4).

3

with Y_{Initial, i} the biomarker measures obtained with the initial assay for an individual i at the M0 visit

4

Y_{Updated, i, j} the biomarker measures obtained with the updated assay for an individual i at visit j = (M24, M48).

These functions were then used to convert measurements from one assay to the other. The latent process model was estimated using the multlcmm function of the lcmm R package [30], independently for each biomarker. The final LPM conversion equations were derived from the estimated parameters of the LPM observation equations as: , where hat denotes the estimate.

LPM were fitted using all biomarker measurements from M0, M24 and 48, excluding the M0 measurements used to derive the bridging equations.

Prediction error

We used the M0 subsample employed to develop the bridging equations as a validation sample, since biomarkers concentrations from both assays were available in this subset. The prediction error was defined as the difference between the observed value (from the updated assay) and the predicted value derived from Eqs. 1 and 2 for bridging and Eqs. 3 and 4 for LPM. Models performance was evaluated with the median absolute error (MAE), the standardized root mean square error (sRMSE) and the coefficient of determination (R²). For the LPM approach, prediction errors were directly computed on the validation sample, as it was not used for model development. An independent validation sample was not available for the bridging approach, then prediction errors were estimated using a LOOCV procedure to minimize bias.

For both approaches, we applied a parametric bootstrap technique to generate 1000 predicted values based on the asymptotic distribution of the parameter vectors from the conversion equations (function mvnorm from the MASS R package). MAE, sRMSE and R² were computed for each of the 1000 prediction sets, and we reported the median values and 95% confidence intervals (95% CI), derived from the 2.5th and 97.5th percentiles.

Biomarker trajectories

We modeled the trajectories of the biomarkers over time using LMM focusing on two variables relevant in the field of AD: the education level (< baccalaureate vs ≥ baccalaureate) and the baseline cognitive impairment (defined by a Clinical Dementia rating score [31] at inclusion at 0.5, vs 0). The effect of each variable on the trajectories was assessed through an interaction term with time (in years), and the regression coefficients obtained from the different conversion approaches were compared. The LMM were adjusted for sex and baseline age, and included random effects on the intercept and time, as well as heteroscedastic measurement errors to account for the differing measurement errors between the updated assay and the converted initial assay.

We conducted three parallel analyses that considered for the M0 biomarker values (i) the raw initial assay measurement (i.e. without conversion) and a binary indicator for the assay (initial/updated) in the regression or, (ii) the conversion derived from the bridging equation or (iii) the LPM equation. To account for error propagation when using the conversion equations in analyses (ii) and (iii), we generated 100 samples of converted measurements using a parametric bootstrap and fitted an LMM to each sample. The final regression coefficients and their 95% CI were estimated by combining the results from the 100 LMMs using Rubin’s rule [32].

Biomarkers were analyzed separately, and systematically transformed with natural logarithm prior to analysis. All statistical analyses were performed using R (version 4.5.0) using RStudio.

Results

Sample characteristics

Among the 2323 individuals enrolled in the MEMENTO cohort, 2299 had at least one blood biomarkers of interest measured at M0, M24 or M48 visit. The median baseline age was 72 years, 62% of participants were female and 55% had an education level of baccalaureate or higher (Table 2). Table 1 summarizes the p-tau181 and NfL quantification for the entire cohort at the three time points, as well as for the bridging and the LTLV studies. The distributions of the biomarkers at follow-up visits and by age are represented in Fig. 2.

Table 2.

Population characteristics. The MEMENTO cohort

	Overall (N = 2299)
Age at study entry, median (q1, q3)	71.6 (65.5, 77.1)
Female, n (%)	1423 (61.9)
Baccalaureate level or above, n (%)	1264 (55.1)
Mild Cognitive Impairment at enrolment (a), n (%)	1358 (59.4)
Presence of at least one APOE eps4 allele, n (%)	658 (30.0)
Follow-up length in months, median (q1, q3)	60.0 (48.8, 61.2)
Time points available for biomarker quantification, n (%)
1	409 (17.8)
2	414 (18.0)
3	1476 (64.2)

APOE Apolipoprotein E

(a) Defined by a Clinical Dementia Rating scale score at 0.5. Other participants were cognitively unimpaired (CDR = 0)

Fig. 2.

Observed blood biomarker distributions. In B, the curves depicting the relationship between the age and the biomarker concentration were obtained by LOESS regression and are presented for descriptive purpose. A Distribution by follow-up visit. B Distribution by age at visit

Conversion equations for NfL and p-tau181

The bridging model based on wDR presented a better fit for p-tau181 (R² = 0.865) compared to NfL (R² = 0.612). For both biomarkers, the OLS model performed slightly better than wDR (Supplementary Table 1, Supplementary Fig. 1). The OLS models exhibited the highest performance for p-tau181 (R² = 0.866) and for NfL (R² = 0.617) (Supplementary Table 1).

For the updated technique, no substantial inter-lot differences were observed for NfL, whereas for p-tau181, the intercept differed significantly from 0 (0.21 [95%CI 0.03;0.39]) (Supplementary Fig. 1, Supplementary Table 3). Consequently, for this biomarker, 0.21 was added to Eq. (1) to account for the lot change.

The links between the latent process and the biomarker measurements are presented in Supplementary Fig. 2 and Supplementary Table 2.

Comparison of approaches

Converted values were similar for the two approaches around the central part of the observed measures, while stronger discrepancies were observed at the distribution tails (Fig. 3). The confidence intervals were wider for the bridging approach than for LPM, as the former was derived from a subsample of M0 measurements, and the latter leveraged data from M0, M24, and M48. The bridging approach demonstrated slightly better performance than LPM for both biomarkers with lower error rates (Fig. 4A and B). After conversion, a shift in the biomarker distribution was observed, aligning it closely with the updated assay (Fig. 5).

Fig. 3.

Comparison of biomarkers conversion according to the bridging and the latent process modelling approaches. Legend: green: conversion through bridging, pink: conversion through latent process modelling. Transparency bands represent confidence interval at 95% obtained by parametric bootstrap on 1000 samples. Histograms represent the distribution of the biomarker measured, solid and dashed lines the median, q1 and q3 respectively

Fig. 4.

Calibration curves in error rates for predicting biomarker concentration. A. Calibration curve. The black diagonal symbolizes the perfect prediction. B. Indicators of prediction error. Error rates were computed for the M0 bridging subsample where measured values were available for both the initial and updated assays. Predictions for bridging were derived from a LOOCV approach. Confidence intervals were computed after the generation of 1000 equation conversions based on the asymptotic distribution of their parameters. LPM: Latent process model, MAE: Median of the absolute error, sRMSE: standardized root of the mean square error

Fig. 5.

Converted and non-converted blood biomarker distributions. A: according to individuals’ age. B: according to the follow-up visit. The curves depicting the relationship between age and biomarker concentration were obtained by LOESS regression and are presented for descriptive purpose. LPM: latent process model

Longitudinal biomarker trajectories

When modeling the trajectories of the biomarkers over the follow-up period using the two conversion approaches or the raw initial measurements with a binary indicator for the assay in the regression, we observed only small differences in the estimated regression parameters (Table 3). These differences did not affect the overall interpretation of the results. Notably, the LPM approach and the binary indicator for assay version yielded very similar results.

Table 3.

Association of education and cognitive impairment on biomarker trajectories according to conversion approaches

	Indicator for assay change		Bridging		LPM
	Beta [95%CI]	p-value	Beta [95%CI]	p-value	Beta [95%CI]	p-value
NfL
Baccalaureate (ref = no)
Effect at baseline	0.042 [0.009; 0.076]	0.012	0.035 [0.007; 0.063]	0.016	0.046 [0.010; 0.082]	0.013
Effect of time (years)	0.025 [0.015; 0.036]	< 10–4	0.024 [0.013; 0.035]	< 10–4	0.026 [0.015; 0.036]	< 10–4
Bac * time	0.001 [− 0.008; 0.01]	0.836	0.001 [− 0.009; 0.012]	0.776	0.000 [− 0.009; 0.010]	0.936
Cognitive impairment (ref = no)
Effect at baseline	0.067 [0.034; 0.101]	< 10–4	0.048 [0.020; 0.077]	0.001	0.070 [0.033; 0.106]	< 10–4
Effect of time (years)	0.021 [0.011; 0.031]	< 10–4	0.018 [0.007; 0.029]	0.002	0.021 [0.010; 0.033]	< 10–4
Cognitive impairment * time	0.008 [− 0.001; 0.018]	0.076	0.012 [0.002; 0.022]	0.019	0.008 [− 0.001; 0.018]	0.092
p-tau181
Baccalaureate (ref = no)
Effect at baseline	−0.002 [− 0.042; 0.039]	0.936	−0.003 [− 0.051; 0.045]	0.915	−0.005 [− 0.049; 0.039]	0.822
Effect of time (years)	0.029 [0.018; 0.040]	< 10–4	0.030 [0.020; 0.041]	< 10–4	0.030 [0.020; 0.040]	< 10–4
Bac * time	−0.002 [− 0.013; 0.009]	0.696	−0.002 [− 0.012; 0.009]	0.715	−0.001 [− 0.011; 0.010]	0.881
Cognitive impairment (ref = no)
Effect at baseline	0.155 [0.111; 0.199]	< 10–4	0.137 [0.099; 0.175]	< 10–4	0.169 [0.125; 0.212]	< 10–4
Effect of time (years)	0.037 [0.026; 0.049]	< 10–4	0.036 [0.026; 0.046]	< 10–4	0.041 [0.031; 0.051]	< 10–4
Cognitive impairment * time	−0.016 [− 0.027; − 0.005]	< 10–4	−0.012 [− 0.022; − 0.002]	0.023	−0.021 [− 0.031; − 0.010]	< 10–4

LPM Latent process model, 95%, CI Confidence interval at 95%

Sensitivity analysis

We replicated this analysis using a bridging conversion based on OLS rather than wDR, as the OLS models demonstrated superior performance in the bridging study for both biomarkers (Supplementary Table 1). Although the prediction error rates for the conversion slightly decreased (Supplementary Table 2, part A), no substantial differences were observed in the final estimation of the LMM coefficients (Supplementary Table 2, part B).

Discussion

In a clinical cohort of 2,299 individuals, aiming at modeling the longitudinal trajectories of blood biomarkers, we compared two methods for managing the impact of changes over time in the measurement assays of two biomarkers. The bridging approach required additional blood sample analyses in a subsample of individuals, whereas the latent process longitudinal model approach utilized all available biomarker measurements. By using age as the time scale, the LPM established links between the two assays and their underlying common process, thereby facilitating conversion from one assay to the other.

The LPM approach exhibited almost comparable performance than bridging for both NfL (R² = 0.49 [95%CI 0.47;0.51] vs 0.55 [95%CI 0.41;0.66]) and p-tau181 (R² = 0.83 [95%CI 0.82;0.85] vs 0.85 [95%CI 0.82;0.88]. The bridging approach led to lower prediction error rates for NfL (MAE = 0.29 [95%CI 0.26;0.33] vs 0.30 [95%CI 0.30;0.31]) and for p-tau181 (MAE = 0.25 [95%CI 0.22;0.28] vs 0.27 [95%CI 0.26;0.28]). After converting M0 measurements using either the bridging or LPM approach, the two variables of interest (education level and cognitive impairment) demonstrated consistent associations with biomarker trajectories. These results were robust, whether the bridging equation was based on weighted Deming or OLS regression.

In long-term follow-up studies, changes in measurement tools are likely to occur. These changes can be explicit, such as introduction of new scanners, updates in assessment tools modalities, or transition between laboratories. Alternatively, they may be subtler, e.g. software updates, device recalibrations, or changes in consumable manufacturers. At the time of quantifying AD biomarkers in MEMENTO, no clinically validated kits were available. Consequently, we relied on assays still under development for research purposes. The results obtained were highly dependent on specific laboratory conditions, rendering the conversion equations derived from this study context-specific and not generalizable to other settings. The bridging and LTLV studies presented here required numerous blood samples from participants, a precious and limited resource that cohorts must allocate judiciously.

When the objective is to model longitudinal trajectories, LPMs offer the advantage to not requiring additional data collection, such as blood samples in our example. Also, this approach relied on limited statistical machinery involving existing tools, and may be applicable in many situations. Unlike bridging studies, LPMs can accommodate multiple changes in measurement tools over time [19], a scenario more likely to arise when analyzing routine data, such as those from hospital data warehouses. Similar to bridging, the conversion equation derived from LPM can be developed once and applied to all subsequent analyses involving the same dataset. However, the equation parameters are study-specific and do not provide a universal conversion between the two techniques. Nevertheless, the LPM approach we proposed can be applied to other datasets to estimate dataset-specific conversion parameters. It is important to note that, while the LPM approach is valuable for statistical purposes, it is not a substitute for laboratories' internal protocols when implementing new tools or updated versions of existing assays in daily clinical practice.

Unlike previous studies [18–20], which modeled trajectories of latent variables, our study demonstrated that conversion to a single measurement tool is feasible. The longitudinal biomarker trajectories are expressed in the natural units of the biomarker (pg/mL), thereby facilitating clinical interpretation and enhancing comparability with other studies. In our case, both the initial and updated assays measured the same quantity in the same unit, and we assumed a linear correspondence between the two assays. In these conditions, bridging, LPM, and regression models applied to raw biomarker data—including an indicator for the assay version—yielded comparable parameter estimates when modeling biomarker trajectories.

This work presents several strengths. First, we examined two distinct biomarkers that underwent different types of assay changes: NfL quantification transitioned from a dedicated assay in 2021 to a multiplex assay in 2023, while the p-tau181 measurement kit evolved from version 2 to version 2.1. Despite these differences, our conclusions remained consistent across both biomarkers. Second, we compared LPM predictions to reference values, an approach that is not typically feasible in latent process analyses. Such analyses usually model unobserved quantities, such as cognition or depression [33, 34], where prediction quality cannot be directly quantified using prediction error metrics, as we were able to do in this study. Finally, our results were robust with an alternative definition of the bridging equation based on OLS regression model.

As a limitation, we acknowledge that the bridging study was conducted a few months after the M24/M48 biomarker quantification, leading to an additional inter-lot validation study. Thus, a correction was required for p-tau181 measurements. Unlike the LPM approach, an independent validation sample for the bridging model was not feasible due to logistic and financial constraints. To address this limitation, we used LOOCV predictions to estimate error rates. We also recognize that our work is limited to the case of a linear correspondence between the two techniques, an assumption that is usually made for laboratory markers. We leave for further research the investigation of non-linear correspondences. Finally, the blood biomarkers analyzed in MEMENTO were measured using Quanterix's SIMOA research assays. These assays are not intended for clinical use, notably due to their variability between kits and lots over time, as we observed here, making our corrections methods necessary. AD biomarkers can now be measured with kits certified for clinical use [35]. While this advancement would facilitate longitudinal follow-up and inter-laboratory comparisons, it does not lessen the importance of our work regarding global applications.

Conclusions

Repeated measurements of disease characteristics or risk factors is the cornerstone of the longitudinal cohort studies. This study illustrates the feasibility and the performance of latent process models for managing changes in biomarker quantification assays when modelling biomarker trajectories over time. The results are presented in the natural units of the biomarkers, facilitating their interpretation and comparability across studies. LPM approach yielded results similar to bridging studies, without requiring dedicated blood sample analysis.

Authors’ contributions

All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by VB, CPL, ABM and MR. The first draft of the manuscript was written by VB and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

The MEMENTO cohort has been funded by the Fondation Plan Alzheimer (Alzheimer Plan 2008‐2012), through the Plan Maladies Neurodégénératives (2014‐2019), and the French Ministry of Research (MESRI, DGRI 2020–2024). This work was also supported by CIC 1401‐EC, Bordeaux University Hospital (CHU Bordeaux, sponsor of the cohort), Inserm, and the University of Bordeaux. The MEMENTO cohort has received funding support from AVID, GE Healthcare, and FUJIREBIO through private–public partnerships. Sponsors and funders were not involved in the study conduct, analysis, and interpretation of data.

Data availability

MEMENTO data access request is available via the Dementia Platform UK Data Access appliance form (https://portal.dementiasplatform.uk/Apply) or via the MEMENTO Secretariat (sophie.lamarque@u-bordeaux.fr).

Declarations

Ethics approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki. The study protocol was approved by the ethics committee “CPP Sud-Ouest et Outre-Mer III”.

All participants enrolled in MEMENTO gave their informed consent.

Competing interests

During the past 3 years, VP was a local unpaid investigator or sub-investigator for clinical trials granted by NovoNordisk, Biogen, TauRx Pharmaceuticals, Janssen, Green Valley Pharmaceuticals and Alector. He received consultant fees for animal studies from Motac Neuroscience Ltd, outside the submitted work. All other authors have no competing interest to declare.