Speech digital biomarker combined with fluid biomarkers predict cognitive impairment through machine learning

Abstract

Background

Current methods for the early detection of Alzheimer’s disease (AD) are constrained by high costs, invasiveness, and limited accessibility, underscoring the urgent need for alternative approaches that are accessible, affordable, and patient-friendly. Previous research has identified speech analysis as a promising tool for the early diagnosis of cognitive impairment (CI). However, the correlation between speech tests and underlying pathology remains undetermined or even obscure. Its clinical utility still lacks pathological validation. We need to further explore the relationship through large-sample analysis and further construct models that can diagnose CIf.

Methods

1223 participants including probable AD or AD (n = 238), amnestic mild cognitive impairment (aMCI) (n = 461) and cognitively unimpaired (CU) (n = 524) were recruited. The participants underwent neuropsychological tests, speech recordings of the “cookie-theft” task, serum biomarker quantification, APOE genotyping, and part of them underwent Aβ PET imaging. Partial Correlation Analysis and LOWESS were used to explore the correlation between speech digital biomarkers and other core AD biomarkers. Finally, machine learning such as XGBoost and Logistic regression were used for constructing the most cost-effective models for CI and Aβ status, leveraging SHAP values for screening.

Results

Significant differences in AD biomarkers were observed among different groups. Notably, the speech digital biomarker percentage of silence duration (PSD) was correlated with cognitive level, serum glial fibrillary acidic protein (GFAP), neurofilament light chain (NFL), phosphorylated tau protein 217 (p-Tau217) and amyloid deposition in specific brain regions. Additionally, we discovered that as the different stages of Aβ deposition progress, PSD, p-Tau217, and GFAP exhibit a two-stage change pattern. Based on the findings, a machine learning CI diagnostic model (AUC = 0.928, 95% CI 0.897 to 0.960) incorporating PSD, APOE, GFAP, and demographic information was developed. Furthermore, an Aβ status classification model (AUC = 0.845, 95% CI 0.783 to 0.907) with PSD, APOE, p-Tau217, and demographic data was also constructed.

Conclusion

Combining speech digital markers with serum and other biomarkers helps identify CI, representing a promising advance in AD detection. This study serves as a preliminary yet encouraging step toward applying speech digital biomarkers in AD diagnostics.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-025-01877-6.

Background

Recently, the increasing focus on early-stage Alzheimer’s disease (AD) diagnosis, spurred by advancements in disease-modifying therapies, has amplified the demand for scalable, accessible biomarkers [1]. Currently accepted diagnostic tools, such as amyloid-beta (Aβ) positron emission tomography (PET) and cerebrospinal fluid (CSF) analyses, encounter limitations in availability, invasiveness, and practicality especially in large-scale screenings [2, 3]. Therefore, these constraints underscore the urgent need for alternative biomarkers to identify patients with cognitive impairment (CI).

As one of the earliest symptoms of AD, speech deficits manifest even prior to discernible memory impairments [4], rendering speech analysis a promising approach for early diagnosis [5–7]. Our previous studies have identified specific speech biomarkers, especially the percentage of silence duration (PSD), which effectively distinguishes early AD or amnestic mild cognitive impairment (aMCI) from cognitively unimpaired individuals (CU) with a high accuracy [8, 9]. A few studies have focused on speech diagnostics and AD pathology indicators such as CSF Aβ and T-tau, but most of them have unstable conclusions due to small sample sizes [10]. Further, diagnostic models that rely solely on one speech modality indicator often lack accuracy and specificity [10, 11].

PSD, as a potential biomarker, has shown promising diagnostic accuracy and cost-effectiveness in preliminary investigations. It overcomes the limitations of cognitive scale assessments, PET scans, and CSF tests, which are associated with substantial medical resource requirements and invasiveness. However, large-scale cohort validation, including serum biomarkers, remains lacking. Here, in the present study, based on the cohorts, we preliminarily explored the correlation of the language abilities represented by PSD with AD biomarkers such as serum biomarkers and AV45 PET. Further, we aimed to develop a model to diagnose CI and further speculate on possible Aβ status by integrating speech features, demographic data, APOE genotyping, and serum markers. We established a novel, accessible framework for CI in the present study, which will aid in the early diagnosis of Alzheimer’s type cognitive impairment with a cost-effective method.

Method

Participant

1223 participants were enrolled, including 461 with aMCI and 238 with probable AD dementia, recruited from the MIMI cohort—a hospital-derived cognitive impairment cohort—as the case groups (Clinical Trials: NCT06534658, registration date: 2024-07-18). Additionally, 524 age-matched CU were recruited from the SheMountain aging cohort (Clinical Trials: NCT05667935, registration date: 2022-12-13) as the control group. Both cohorts were recruited based on the same criteria and participants were recruited as early as 2022 from Shanghai. All procedures contributing to this work adhere to the ethical standards set by the relevant national and institutional committees on human experimentation, as well as the Helsinki Declaration of 1975, as amended in 2008. The Ethics Committee of Ruijin Hospital approved all procedures involving human participants/patients (approval number: 2022 - 254). Written consent was obtained from all included participants. A flowchart detailing the enrollment of participants and the research methods employed in this study is available in Supplementary Fig. 1.

Clinical assessment and diagnosis

All participants were required to provide demographics (age, gender, years of education), family history, and medical history, and underwent neuropsychological tests including the following: the Montreal cognitive assessment (MoCA), the Clinical Dementia Rating scale (CDR) and the Cookie-theft picture description task from the Boston Diagnostic Aphasia Scales.

These diagnoses were based on cognitive tests after recruitment. Clinical diagnoses of MCI and probable AD or AD were based on the NIA-AA criteria (2011) [12]. In the following text, we will refer to it as AD uniformly. CU group were considered as cognitively unimpaired after the clinical consultation. Participants were excluded if they had cognitive impairment due to other degenerative diseases such as frontotemporal lobar degeneration, vascular dementia, Lewy body dementia, etc., any systemic disease which can lead to cognitive dysfunction, psychiatric disorders, or severe hearing or vision impairment.

Speech recording protocol

All participants performed a Cookie-Theft picture description task [13], during which they were given a picture and were told to discuss everything they could see happening in the picture in 1 min while being recorded. The speech was recorded under the following configuration parameters of Cool Edit Pro software: a frequency of 160,000 Hz, creating a 16-bit mono recording, and environmental noise was limited to under 45 dB. The automatic speech recognition (ASR) software for CI v1.3 (China Software Copyright number 2016SR164680) for speech analysis was used to extract the speech/silence parameters. PSD is defined as the sum of all silent periods divided by the total speech time (ratio of total silent pause duration to total speech duration), expressed as a percentage [8, 9].

APOE genotyping

Genomic DNA from blood samples was purified using the TIANamp Blood DNA Kit (DP348-03, TIANGEN, Beijing, China). APOE genotype was classified by real-time quantitative PCR (qPCR) with genotyping assays (TaqMan, 4351370, USA) and The LightCycler^® 480 System (Roche, USA). The polymorphisms distinguish the ɛ2 allele from the ɛ3 and ɛ4 alleles at amino acid position 158 (rs7412) and the ɛ4 allele from the ɛ2 and ɛ3 alleles at amino acid position 112 (rs429358).

Serum protein quantification

Venous blood was collected in serum tubes (BD Vacutainer^® Serum Tubes, USA) and centrifuged at 2,000 × g for 10 min at 4 °C. The obtained serum was divided into 500 µl aliquots and frozen under − 80 °C for further measurement. All samples underwent no more than three freeze-thaw cycles. When measuring, the samples were rapidly thawed at room temperature and then centrifuged at 10,000 × g for 15 min prior to analysis to prevent any sample debris from interfering with the measurements. Concentrations of serum Glial Fibrillary Acidic Protein (GFAP), Neurofilament Light Chain (NFL), β Amyloid 40 (Aβ40), β Amyloid 42 (Aβ42), Phosphorylated tau protein 181 (p-Tau181) and Phosphorylated tau protein 217 (p-Tau217) were measured using the Single Molecular Immunity Detection (SMID) platform (AST-Sc-Lite; ASTRABIO, China), which is a fully-automated SMID principle-based machine. The samples were measured using a two-step immune process respectively for each biomarker. Briefly, 25 µL serum was quickly mixed with 25 µL Reagent 1, which was mainly composed of capture antibody coated magnetic beads. After 6 min incubation under 40 °C, 10 µL Reagent 2, who containing detection antibody conjugated single-molecule-imaging fluorophore was added following by another quick mix and 4 min incubation. The mixture after reaction was subsequently moved to measuring flow-cell for magnetic beads gathering, forming single-layer beads array, washing and imaging. The concentration of each biomarker was then calculated with pre-prepared standard curve. All analytical procedures were performed according to the protocol of the assay and machine manufacturer by well-trained technicians who were blinded to the state of participant and clinical data, according to the protocol of the manufacturer. Samples with test values exceeding the mean ± 3 times the standard deviation were excluded.

¹⁸F-Flobetapir (AV-45) PET imaging

201 of the participants accepted ¹⁸F-florbetapir (AV45) PET/MR scanning within 2 weeks. PET and MRI were performed with a 3.0-T hybrid PET/MR scanner (Biograph mMR; Siemens Healthcare, Erlangen, Germany) using the vendor supplied 12-channel phase-array head coil. All the participants received an intravenous injection of AV45 at a mean dose of 3.7 MBq/kg body weight to image Aβ, and statistic PET images were acquired at 40–60 min after the bolus injection. The results were jointly obtained by two experienced nuclear medicine physicians. The Standardized Uptake Value Ratio (SUVR) based on Cerebellar value of each brain region were extracted using toolbox SNBPI [14]. The whole brain Aβ burden was estimated by an Aβ meta regions of interest (meta-ROI) mask which includes the frontal, anterior/posterior cingulate, lateral parietal, and lateral temporal regions, and the reference region is the cerebellum [15], which can be used as a template to provide a standardized assessment of Aβ deposition in various brain regions.

Statistical analysis

We excluded extreme values of each digital and serum biomarker, which were those at least three times the SD away from the mean. For the handling of missing data, we adopted the method of imputing with the mean value within the diagnostic group. The chi-square test was used to analyze differences in categorical variables among groups. ANOVA and Tukey post hoc comparisons were used in continuous variables such as fluid biomarkers, MRI and PET measures to analyze differences among groups. Correlations between PSD, serum biomarkers, and scale test scores were tested using partial correlation analysis, in which age, gender, education level, and APOE genotype were used as control variables. LOWESS was used for the association between PSD, serum p-Tau217, Aβ42/40 and brain amyloid deposition.

Due to varying sample sizes, in order to achieve maximum efficacy while avoiding overfitting, the classification model was constructed using the XGBoost algorithm and logistic regression model, aiming to achieve clinically translatable predictive performance (expected CI diagnostic AUC ≥ 0.90, Aβ prediction AUC ≥ 0.85). For validation of effectiveness, the XGBoost model employed an internal 8:2 training and validation set split, while the logistic regression model, due to sample size limitations, utilized 5-fold cross-validation to ensure model performance closely approached the expected targets. First, we established the best model included demographic, genes, fluid biomarkers, and digital markers. Feature selection was performed exclusively on the training set via (SHapley Additive exPlanations v0.44.0) SHAP-guided backward elimination, ensuring no test data leakage throughout the process. SHAP [16] was also used to quantify each feature’s impact on the model’s output. Finally, the area under the receiver-operating characteristic (ROC) curves (AUC), accuracy, specificity, and the sensitivity were used to evaluate the effect of the model. The difference in the AUC was determined using DeLong statistics. The model with the least parameters and AUC without significantly difference from the best model were selected as the most efficient model. The hyperparameters were optimized using grid search. For the XGBoost model, our optimal hyperparameters were nrounds = 200, max_depth = 6, eta = 0.01, gamma = 0.1, colsample_bytree = 0.8, min_child_weight = 1, subsample = 0.8. As for the logistic regression, the hyperparameters were penalty alpha = 1 (L1 regularization), regularization strength lambda = 0.0285. Due to the sample size, we opted for a more robust logistic regression to prevent overfitting while maintaining high efficiency. Meanwhile, we evaluated the clinical application effects of different models by plotting decision curves through RMDA package. P-values of < 0.05 were considered statistically significant. Given the large-scale population sample, we assumed these continuous variables follow a normal distribution. R version 4.3.2 was used for all statistical analyses.

Results

Demographics, clinical characteristics, core AD fluid, digital and imaging biomarkers among groups

A total of 1223 participants were included in this study, consisting of 461 with aMCI, 238 with AD dementia, and 524 CU. The overall proportion of females was 57.40%, with the highest proportion in the AD group (68.91%) and the lowest in the CU group (48.90%). The average age was 74.06 ± 7.28 years, with no significant differences in age among groups (F = 2.9267, P = 0.0539). The CU group had the longest educational years (9.52 ± 4.51 years), while the MCI and AD groups had shorter educational years. APOE-ε4 carriers accounted for 23.96%, with the highest carrier rate in the AD group (37.39%) and the lowest in the CU group (19.27%) with significant differences (X² = 30.56, P < 0.0001). The detail and post hoc analysis can be found in the Table 1.

Table 1.

Demographics of participants

Variable	Total (n = 1223)	CU (n = 524)	MCI (n = 461)	AD (n = 238)	Stat_value	P_value
Gender, Female, (%)	702/1223(57.40)	256/524(48.9)	282/461(61.17)	164/238(68.91)	X2 = 31.2200	< 0.0001
Age, mean (SD), years	74.06(7.28)	73.51(6.49)	74.62(6.73)	74.20(9.55)	F = 2.9267	0.0539
Education, mean (SD), years	8.30(4.21)	9.52(4.51)	7.30(2.39)a	7.56(5.45)a	F = 41.0725	< 0.0001
APOE-ε4 carriers, n (%)	293/1223(23.96)	101/524(19.27)	103/461(22.34)	89/238(37.39)	X2 = 30.5600	< 0.0001
MoCA, mean (SD)	17.70(7.85)	24.40(3.63)	14.77(5.62)a	8.61 (5.39)a, b	F = 1028.9338	< 0.0001
CDR, mean (SD)	0.48(0.61)	0.01(0.06)	0.50(0.02)a	1.44(0.69)a, b	F = 1759.7469	< 0.0001
GFAP, mean (SD), pg/ml	8.13(5.13)	6.00(3.33)	9.14(5.21)a	10.89(6.23)a, b	F = 103.3351	< 0.0001
NFL, mean (SD), pg/ml	11.40(6.63)	10.23(5.52)	11.49(5.47)a	13.83(9.62)a, b	F = 25.1172	< 0.0001
Aβ40, mean (SD), pg/ml	609.15(257.14)	617.64(271.47)	607.30(221.74)	594.02(287.20)	F = 0.7091	0.4923
Aβ42, mean (SD), pg/ml	69.66(51.62)	74.01(50.28)	67.47(51.07)	64.31(54.95)a	F = 3.5698	0.0285
p-Tau181, mean (SD), pg/ml	1.56(0.97)	1.43(0.72)	1.63(0.97)a	1.71(1.33)a	F = 8.9750	0.0001
p-Tau217, mean (SD), pg/ml	1.65(1.12)	1.21(0.80)	1.84(1.18)a	2.26(1.20)a, b	F = 95.6686	< 0.0001
Aβ42/40, mean (SD)	0.19(0.93)	0.25(1.40)	0.14(0.16)	0.17(0.25)	F = 2.0699	0.1266
PSD, mean (SD)	0.39(0.34)	0.23(0.14)	0.51(0.37)a	0.53(0.42)a	F = 130.8601	< 0.0001
AV 45-PET, positive/total	152/201	23/72	21/21	108/108	-	-

^aSignificant values versus CU

^bSignificant values versus MCI

SD, Standard Deviation

The digital and fluid biomarkers were significantly different among the three groups. GFAP (F = 103.3351, P < 0.0001), NFL (F = 25.1172, P < 0.0001), p-Tau181 (F = 8.9750, P = 0.0001), p-Tau217 (F = 95.6686, P < 0.0001), these four indicators were the lowest in the CU group and the highest in the AD group, and there were significant differences between the groups. On the contrary, Aβ42 and Aβ42/40 showed a potential lower level in MCI or AD groups than that in CU group without significant differences (F = 2.0699, P = 0.1266). Regarding PSD, the digital biomarker, gradually increased with the course of the disease, and PSD in AD group has the biggest value against that in the CU group with a significant difference (F = 130.8601, P < 0.0001). The details and post hoc analysis can be found in the Table 1.

Since the assessment of amyloid pathology is crucial in the ATN framework of AD, AV45-PET examinations were completed in 201 people including 72 CU, 21 aMCI, 108 AD, of which 49 were negative and 152 were positive after visual judgment. The demographic information and AD-related biomarker can be found in the Supplementary Table 1. We further calculated the SUVR of different ROIs. After quality inspection, preprocessing and SNBPI toolbox calculation, we included 123 samples for statistical analysis, of which 37 were negative and 86 were positive. The SUVR in each ROI which had significant associations were shown in the Table 2. As expected, Aβ SUVR in these regions, including frontal lobe (t=-7.0402, p < 0.0001), parietal lobe (t=-7.1796, p < 0.0001), temporal lobe (t=-7.7205, p < 0.0001), occipital lobe (t=-7.7586, p < 0.0001), insula (t=-7.287, p < 0.0001), limbic system (t=-6.4329, p < 0.0001), subcortical (t=-3.7746, p = 0.0003), and total brain (t=-9.2686, p < 0.0001), were significantly higher in the positive group.

Table 2.

Information of AV45 PET imaging

Variable	AV45 negtive (n = 37)	AV45 positive (n = 86)	t Value	P_Value
PSD, mean (SD), %	0.22(0.13)	0.34(0.29)	-4.1194	0.0001
Aβ SUVR in frontal lobe	1.05(0.14)	1.28(0.20)	-7.0402	< 0.0001
Aβ SUVR in parietal lobe	1.05(0.16)	1.29(0.20)	-7.1796	< 0.0001
Aβ SUVR in temporal lobe	1.01(0.13)	1.23(0.18)	-7.7205	< 0.0001
Aβ SUVR in occipital lobe	1.20(0.12)	1.42(0.18)	-7.7586	< 0.0001
Aβ SUVR in insula	1.01(0.11)	1.22(0.21)	-7.2870	< 0.0001
Aβ SUVR in limbic system	1.14(0.11)	1.32(0.20)	-6.4329	< 0.0001
Aβ SUVR in subcortical	1.12(0.12)	1.23(0.20)	-3.7746	0.0003
Total brain Aβ SUVR	0.98(0.14)	1.27(0.20)	-9.2686	< 0.0001

PSD, percentage of silence duration

SD, Standard Deviation

Correlation of PSD with cognitive assessments, fluid biomarkers, and amyloid burden in the brain

To further investigate the correlation between PSD and AD-related markers, the significant associations between them from 3 aspects were found after controlling for age, gender, education level, and APOE genotype: cognitive level, fluid biomarkers, and amyloid burden. ⑴ Correlation with cognitive level: PSD was negatively correlated with MoCA (r = -0.3285, P < 0.0001) (Fig. 1B), and showed significant differences between groups, with the smallest value in the CDR 0 group (P < 0.0001) (Fig. 1B); ⑵ Correlations with serum biomarker: PSD was positively correlated with GFAP (r = 0.0737, P = 0.0108), NFL (r = 0.0603, P = 0.0372) and p-Tau217 (r = 0.1208, P = < 0.0001), rather than Aβ42/40 (r = -0.0357, P = 0.2172)(Fig. 1C-F); ⑶Correlation with AV45 SUVR: PSD had no significant correlation with whole brain amyloid burden (r = 0.1779, P = 0.0530). But in further analysis of brain regions, there were positive correlation with AV45 SUVR in the frontal lobe (r = 0.2042, P = 0.0259), temporal lobe (r = 0.2111, P = 0.0212), except the whole brain burden presented as Aβmeta SUVR(r = 0.1779 P = 0.053) (Fig. 2A-C). Other partial correlation analysis can be found in Supplementary Table 2.

Fig. 1.

Correlation of PSD with cognitive assessments, fluid biomarkers. (a) Pearson correlation curve between PSD and Moca; (b) Differences in PSD under different CDR scores; (c-f) Pearson correlation curve between PSD and different serum biomarkers (Aβ42/40, GFAP, NFL, p-Tau217) and the r value, p value of partial correlation analysis

Fig. 2.

Correlation of PSD with amyloid burden in the brain, and the trend of biomarkers at different Aβ deposition levels. (a-c) Pearson correlation curve between PSD and different AV45 PET ROI SUVR (frontal lobe, temporal lobe, and the AV45 meta SUVR) and the r value, p value of partial correlation analysis

The speech digital biomarkers and serum biomarkers are affected by central amyloid pathology. We used the LOWESS model to measure the association between speech digital biomarkers, serum p-Tau217, and Aβ42/40, and central amyloid deposition. Surprisingly, the results of correlation analysis revealed that PSD and serum p-Tau217 were sensitive to amyloid deposition, which appeared to have two distinct stages. In the stage before the inflection point of amyloid deposition changes (Aβ meta SUVR < 1.20), PSD, serum, p-Tau217 and GFAP showed an upward trend, but after the inflection point, they reached a plateau and no significant increase is observed (Fig. 3A-C). However, the correlation of Aβ42/40 with amyloid deposition was not significant (Fig. 3D) with an upward trend.

Fig. 3.

The trend of biomarkers at different Aβ deposition levels. (a-d) LOWESS curve between AV45 meta SUVR and different biomarkers (PSD, p-Tau217, GFAP, Aβ42/40, NFL) and the corresponding r value, p value of Pearson correlation

Combining digital and fluid markers to identify individuals with cognitive impairment

Demographic information, speech digital biomarkers, serum biomarkers, and genetic information (APOE), were used to identify CI. We used 1223 participants as modeling samples, with aMCI, probable AD or AD as the CI population as the identification target. The best model containing all features was established using the XGBoost model with an AUC of 0.951 (95%CI, 0.925 to 0.976). Finally, through excluding the low-weight serum biomarkers by SHAP values (Supplementary Fig. 2), we found the simplest and most efficient model, whose prediction AUC in the test set was 0.928 (95%CI, 0.897 to 0.960), and there was no significant difference between the AUC and the best model. The model included age, sex, years of education, APOE genotype, PSD, GFAP (Fig. 4A). PSD contributed the most among all parameters, while GFAP is the largest among all serum indicators. The screening process of the model can be found in Supplementary Table 3. Through SHAP values, we can observe that PSD and GFAP contribute the most in this model, and their interaction also has a significant impact on the model. The SHAP importance of each feature in the model can be found in Supplementary Fig. 2B, D-I. The decision curves for each model are shown in Supplementary Fig. 2 J.

Fig. 4.

The diagnosis model of cognitive impairment screen and prediction of amyloid status in the brain. (a, b) The XGBoost and Logistic regression model process. The best fit model shows the data-driven model selection with the highest area under the receiver-operating characteristic (ROC) curve (AUC). In subsequent models, modalities were removed step by step to obtain a similar and efficient model performance with as few significant measures as possible. Comparisons between AUCs were performed using DeLong statistics. (a, right) ROC curve analyses of the different models for discriminating cognitive impairment. (b, right) ROC curve analyses of the different models for discriminating Aβ status as determined by positron emission tomography (PET) scans.PSD, percentage of silence duration; GFAP, Glial Fibrillary Acidic Protein; NFL, Neurofilament; 95%CI, 95% Confidence Interval; *, P < 0.05, **, P < 0.01

Combining digital and fluid markers to speculate on possible Aβ status

Next, we tried to assess whether combining digital with serum biomarkers could further improve the accuracy of predicting PET Aβ status. We selected 201 participants who completed AV45 PET imaging as the sample, of which 152 amyloid-positive patients served as the identification target. With the same method, we constructed the most efficient Logistic regression model with AUC = 0.845 (95% CI, 0.783 to 0.907). The model included age, gender, years of education, APOE genotype, PSD, and p-Tau217 (Fig. 4B). Serum p-Tau217 contributed the most among all parameters. The screening process of the model can be found in Supplementary Table 4. Through SHAP values, we can observe that p-Tau217, APOE genotype, and PSD contribute the most to this model. The SHAP importance of each feature in the model can be found in Supplementary Fig. 3B-H. The decision curves for each model are shown in Supplementary Fig. 3I.

Discussion

In the present study, we found that: (1) PSD is correlated with AD-related peripheral blood markers, especially serum GFAP and p-Tau217. (2) PSD is correlated with amyloid burden in the brains regions: the frontal lobe and temporal lobe. (3) PSD and serum p-Tau217 were sensitive to amyloid deposition which appeared to have two distinct stages. (4) After combining demographic information, genetic susceptibility gene APOE information, speech digital markers and serum biomarkers and other features, the model can well identify people with CI and might further speculate on possible Aβ status. In summary, it provides a feasible and novel scheme for CI screeningβ.

As a symptom that can appear early in AD, speech impairment may be caused by structural or functional damage resulting from AD-related pathologies. We reported positive results between PSD with GFAP, NFL, and p-Tau217 for the first time. However, we didn’t find the significant correlation between serum Aβ42 and PSD. But a significant correlation between CSF Aβ42 and lexical-semantic scores has been found in a small sample study mention above [10]. The main reason may be the different samples types and detection methods may have caused this reason, that SMID and single molecule array (Simoa) may lack the sensitivity to accurately measure total serum Aβ levels compared to mass spectrometry-based methods [17]. Generally, serum GFAP [18] reflected neuroinflammation and permeability of the blood-brain barrier, NfL [19] the biomarker of neuro-axonal injury, p-Tau217 [20, 21] all have a good correlation with AD cognitive level, and are widely used in the construction of diagnostic models [17, 22]. Therefore, we hypothesize that language function impairment may be associated with the aforementioned AD pathological processes. However, this potential link requires further validation through longitudinal studies and targeted research methodologies. Meanwhile, the correlation between PSD and core AD biomarkers further supports its utility as a biomarker for assessing cognitive impairmentβ.

We found the frontal lobe, temporal lobe positively correlated with PSD in the present study. But we can’t find significant correlation between the PSD and the whole brain Aβ burden defined by Aβ meta-ROI mask, which may be caused by the sample size. The correlation has also been verified in a small-sample CSF and speech study [10]. Interestingly, as amyloid protein accumulates continuously, PSD, serum p-Tau217 and GFAP do not change linearly, which appeared to have two distinct stages. This inflection point is SUVR = 1.2, which is just slightly higher than the AV45 positive threshold of 1.11 in other PET studies [23]. However, the current findings are only based on cross-sectional data. Longitudinal data will be more convincing. But the present pattern is quite similar to the classic model of the sequence of biomarker changes with disease progression [24]. We speculate in the late stage of amyloid pathology, the downstream tau hyperphosphorylation and neurotoxicity may tend to be stable, and the Aβ monomers or oligomers with obvious cytotoxicity [25] decrease after aggregating into plaques which may be a possible reason. This phenomenon also proves the sensitivity of PSD, serum p-Tau217 and GFAP to changes in the early stage of amyloid protein, so they can be suitable indicators for AD early diagnosis.

The diagnosis of cognitive impairment relied on cognitive scales and exams by skilled physicians, and further PET imaging or CSF test, which are invasive and expensive. Then, we constructed two classification models as following: For predicting CI, the simplest and most efficient model were finally constructed with an AUC of 0.928 (95%CI, 0.897 to 0.960); For speculating on Aβ status, the final model AUC reached 0.845 (95% CI, 0.783 to 0.907), which is close to previous studies using either: [1] demographic data + APOE + Aβ42/40 + p-tau181 + GFAP + NfL to distinguish possible Aβ status [22], or [2] NLP-based CI diagnosis with examiner-evaluated recordings + APOE + demographic data (AUC = 0.926) [6]. Compared to these approaches, our study innovatively combines demographic data, speech features, APOE, and a single serum biomarker, offering distinct advantages in both personnel requirements and cost efficiency. Notably, the inclusion of cognitively unimpaired participants with positive amyloid in the CU group strengthens the model’s credibility. Nevertheless, further clarification through larger sample sizes and longitudinal studies remains necessary.

Interestingly, we used two different optimal serum biomarkers in the 2 models above, GFAP in the first model while p-Tau217 in the second model. This data-driven result is highly consistent with previous studies. GFAP has good sensitivity but poor specificity and can be elevated in other neurological diseases such as stroke [26], Lewy body dementia [27]. P-Tau217 has the relatively best predictive efficiency among these serum biomarkers, and p-Tau217 can predict the longitudinal changes in brain Aβ load in the preclinical AD stage [28]. As for the PSD the speech digital biomarker, it has an important weight in the both 2 models (relatively 1st and 4th in all features according to SHAP values Supplementary Fig. 2, 3). As a tool easily accessible, PSD can reflect cognitive function in the early stage and can well predict the amyloid status in cooperation with other biomarkers. Recently Samad and colleagues [5] predicted MCI-to-AD progression within 6 years accurately, which indicates speech biomarker potential in disease progression. And compared to the simple digital biomarkers, our model integrates multimodal data such as blood and APOE, thus providing a more comprehensive perspective for AD diagnosis. In general, our study provides a novel, easily accessible and cost-effective method to predict CI, which may contribute to large-scale cognitive screening and medication evaluation in areas with limited medical conditions and continuously increasing elderly people.

By integrating non-invasive speech biomarkers with efficient blood-based biomarkers, we enhance diagnostic accuracy and streamline operations. This approach offers a cost-effective alternative to high-resource methods like PET scans and CSF testing. However, our study also has some limitations: (1) This is a single-center clinical cohort. The recruited individuals are from Shanghai, China. Now, multicenter studies and extensive collaborative efforts should be carried out, which would provide a broader validation of the effectiveness of this approach. Additionally, our study has limitations in sample size, particularly regarding participants who have undergone amyloid PET imaging. Consequently, while our current findings show promise, they require validation through larger-scale studies with more comprehensive biomarker data. (2) In this study, participants were primarily recruited from individuals with cognitively unimpaired (CU) and Alzheimer’ s disease (AD)-related cognitive impairment. Given that speech disorders in other dementia types (e.g., frontotemporal degeneration and vascular dementia) exhibit distinct pathological features compared to AD, these cases were excluded. However, our second model does not perfectly predict Aβ status, which may require validation in larger cohorts and longitudinal studies. (3) This study included research on only Aβ PET imaging rather than Tau PET. This aspect will be further explored and refined in subsequent research. (4) This is a cross-sectional study. Further follow up should be performed to verify its effect to predict the disease conversion through longitudinal data. (5) Due to the need for standardized speech collection protocols, our current analysis does not support dialects or noisy environments. Future improvements will focus on enhancing speech analysis software compatibility and implementing standardized voice collection procedures.

Conclusions

Recent advances in digital technology and artificial intelligence have facilitated the use of digital cognitive testing. Our research initially uses speech digital indicators combined with other easily accessible data to predict CI with cost-effectivity. However, future multicenter and longitudinal studies are needed for verification to establish a standardized scheme widely applicable to primary care institutions and promote the development of AD clinical drugs and the diagnosis of AD.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AD

Alzheimer’s disease

CI

Cognitive impairment

aMCI

Amnestic mild cognitive impairment

CU

Cognitively unimpaired individuals

PSD

Percentage of silence duration

GFAP

Glial fibrillary acidic protein

NFL

Neurofilament light chain

p-Tau217

Phosphorylated tau protein 217

Aβ

Amyloid-beta

PET

Positron emission tomography

CSF

Cerebrospinal fluid

MoCA

Montreal cognitive assessment

CDR

Clinical dementia rating scale

APOE

Apolipoprotein E

SMID

Single molecular immunity detection

SUVR

Standardized uptake value ratio

ROI

Region of interest

XGBoost

EXtreme gradient boosting

ROC

Receiver-operating characteristic

AUC

Area under the curve

SHAP

Shapley additive explanations

FTD

Frontotemporal degeneration

AV45

18 F - Florbetapir

Author contributions

Concept and design: JTW, NZ, GX, RJR, GW. Acquisition, analysis, or interpretation of data: JTW, NZ, GX, JLG, JWX, Hai-Xia Li, JPL, XYX, YYS, WWC, RJR, GW. Drafting of the manuscript: JTW, NZ, GX. Critical review of the manuscript for important intellectual content: NZ, GX, JLG, RJR, GW. Administrative, technical, or material support: JTW, GX, RJR, GW. Supervision: RJR, GW.

Funding

This work was supported by Brain Science and Brain-Like Intelligence Technology of the Ministry of Science and Technology of the People’s Republic of China (2021ZD0201804).

Data availability

All data can be accessed after obtaining the author’s consent.

Declarations

Ethics approval and consent to participate

The Ethics Committee of Ruijin Hospital approved all procedures involving human participants/patients (approval number: 2022 − 254).

Consent statement

All participants provided written informed consent.

Consent for publication

All authors have agreed to the publication.

Competing interests

The authors declare no competing interests.