The Kawasaki Aging Well-being Project (KAWP), carried out in Kanagawa, Japan, examines the overall health status and associations with disability in disability-free community-dwelling older adults aged 85–89 years [9]. During the first phase of the KAWP, 1,026 participants went to local facilities in Kawasaki City from March 2017 to December 2018. For the discovery phase of the present analysis, participants who had never needed to be enrolled in the Long-Term Care system (Care level 1 or higher) at the time of study recruitment were chosen at random. Among these participants, approximately one-fifth with complete data on potential covariates (see Other Variables) were randomly selected for plasma protein measurements owing to the resource and cost constraints inherent to multiplexed proteomic assays. We used machine learning-based variable selection; thus, conventional sample size calculations were not performed owing to the lack of prior knowledge on effect sizes, consistent with previous studies [8]. However, for reference, we estimated the required sample size assuming a conventional Cox model with cystatin C. Based on a previous study [7], assuming alpha of 0.05, beta of 0.10, and event rates of 20–30%, the required sample sizes ranged from 86 to 129 participants (R package, powerSurvEpi v.0.1.3). All participants received an extensive explanation of the study and participated after providing written informed consent. The study protocol conformed to the principles of the Declaration of Helsinki, with the approval of the Ethics Committee of Keio University School of Medicine (ID: 20,160,297). The KAWP study was also registered in the University Hospital Medical Information Network Clinical Trial Registry as an observational study (ID: UMIN000026053).

A total of 38 plasma proteins were quantified on samples obtained from the discovery cohort using the Luminex® xMAP® technology (Supplementary Table S1). The samples were incubated for an entire night in plates at 4˚C, with fluorescent-coded magnetic beads, each conjugated with antibodies against the analyte of interest. The plates were then washed and biotinylated antibodies were incubated using the complex. Following incubation, the plates were washed, and Streptavidin-PE was added to each well and incubated. The plates were washed and resuspended in sheath fluid for measurement on the Luminex™ FLEXMAP 3D®. Data were collected using xPONENT® 4.0 (Luminex) software and assessed using Bio-Plex Manager™ Software 6.1.1 (Bio-Rad). Sample concentrations were calculated using a 4PL or 5PL algorithm, depending on the MFI generated by each standard.

We exempted nine proteins from the analysis because more than 40% of the samples fell below the limit of detection (LOD). For the remaining proteins, we replaced values less than LOD with half the minimum value.

This study used registry data on LTCI system certification and vital status from Kawasaki city and Kanagawa prefecture. Information about the LTCI is described in Supplementary Table S2. Incident disability was defined as the first certification of Care level 1 (Yo-kaigo 1) or higher, a classification in the LTCI. Certifications of LTCI system, as well as death, were registered at the registry office corresponding to the municipality where the decedent lived following the occurrence of these events. The LTCI system had a participant follow-up rate (and data availability) for over 5.5 years of 100% in the discovery phase subcohort and 98% in the full KAWP.

We conducted face-to-face interviews to assess participants’ general characteristics, including age, sex, education level (high school and higher education or not), smoking status (never smoker or not), and alcohol drinking habits (never drinker or not). Body mass index (BMI; kg/m2) was calculated by weighing participants and measuring their heights. We calculated the glomerular filtration rate (eGFR) by measuring serum creatinine using a standard enzymatic method as follows: eGFR (mL/min/1.73 m2) = 194* [serum creatinine (mg/dL)]−1.094*[age (years)]−0.287[*0.739 (if female)] [10].

We compared deceased and censored individuals, as well as disabled and censored individuals, using the Student’s t-test for continuous variables and the chi-square test for categorical variables.

Prior to modeling, protein values falling below the LOD were imputed as described in Plasma proteins, and only participants with complete covariate data were retained for the subsequent analyses. To investigate candidate proteins associated with the risks of disability and mortality, we fitted elastic net-regularized Cox proportional hazards models (R package, eNetXplorer v. 1.1.3) incorporating the 29 plasma proteins and adjusting for age, sex, BMI, education, smoking/drinking status, and eGFR [11]. The model explores the best-fitted model by varying the mixing parameter alpha between 0 (Ridge) and 1 (Lasso) [12]; alpha was increased in 0.1-step increments. For each alpha, the best shrinkage penalty parameter (lambda) was chosen using a cross-validation procedure that optimized the out-of-bag correlation between observed and predicted outcomes. For this cross-validation procedure, we ran 1,000 iterations, randomly selecting 80% of samples as the training set and 20% as the validation set for each iteration. We chose the parameter alpha with the highest average correlation between observed and predicted results. To estimate the significance of model coefficients, eNetXplorer generates null models by randomly permuting sample labels 250 times during each cross-validation run. Proteins found in the feature selection phase (using a lenient significance threshold of p < 0.10) were tested using multivariable Cox hazard models, adjusting for the same covariates as in the previous step. A threshold of p < 0.10 was applied as a conservative, hypothesis-generating criterion, consistent with prior proteomic discovery studies [8]. This approach yields more interpretable effect estimates and confirms the strength of the relationships between these proteins and the outcomes outside of the regularization framework. Furthermore, we developed a composite score to evaluate the combined effect of selected proteins on the risk of outcomes. The Z-scores of each protein were added together, with the signs adjusted to reflect their respective directions of association with the results.

To externally examine our results, we used the data from the Invecchiare in Chianti, “Aging in the Chianti Area” (InCHIANTI) [13]. A total of 954 participants had plasma proteomics data determined by the 1.3 K SomaScan assay (SomaLogic, Boulder, CO) [14,15,16], disability assessed by Katz ADL scale and subjective 400 m walk capacity [17, 18], and vital status acquired from the Mortality General Registry data operated by the Tuscany Region and death certificates deposited after death at the Registry Office in the municipality of residence. We used the Cox hazards model to investigate the relationships between the selected proteins in the discovery phase and disability and mortality risks after controlling for the same set of covariates. We also conducted the same analysis on a subset of participants aged 80 years and older. Of note, the InCHIANTI cohort differed substantially from the KAWP discovery sample with respect to age distribution (mean age, 66.1 vs. 87.1 years), follow-up duration (15 vs. 4.5 years), disability assessment methodology (Katz Activities of Daily Living and 400-m walk capacity vs. LTCI certification), and proteomic platform (SomaScan 1.3 K vs. Luminex xMAP).

During the external replication phase, we broadened the analysis to the full KAWP, using plasma B2M, cystatin C, NT-proBNP, and fetuin-A measured with orthogonal experimental techniques and their association with disability risk. To assess the technical reproducibility of measurements, we calculated Spearman’s correlations for each of the four proteins, comparing Luminex measurements to those obtained using these orthogonal methods.

The proportional hazards assumption was verified for all Cox models using Schoenfeld residual tests (cox.zph in R). In both the KAWP discovery and full-cohort analyses, the global tests were non-significant (all p > 0.18). In the InCHIANTI mortality analyses, mild violations of the proportional hazards assumption were observed for the global test of B2M (p = 0.009), cystatin C (p = 0.014), and epidermal growth factor (EGF; p = 0.010). Inspection of the individual Schoenfeld residual terms indicated that these violations were driven by age (p < 10−8) rather than by the protein exposure variables (all protein-specific p > 0.09). No violations were observed in any disability analysis (Supplementary Table S3).

P-values below 0.05 were deemed statistically significant. All statistical analyses were carried out using R v.4.3.1.

All participants gave written informed consent. The study protocol followed the principles of the Declaration of Helsinki with the authorization of the Ethics Committee of relevant institutional Ethics Committee (ID: 20,160,297).