Individualized prognosis of cognitive decline and dementia in mild cognitive impairment based on plasma biomarker combinations

Study population characteristics (model selection sample)

A total of 148 patients with MCI from the Biomarkers for Identifying Neurodegenerative Disorders Early and Reliably (BioFINDER) cohort for whom all three plasma and CSF biomarker measurements were available, along with at least one of the primary or secondary outcomes, were used for model selection (Table 1). The mean age was 71.4 years, the mean education duration was 11.2 years, 36.5% were female and the mean mini-mental state examination (MMSE) score was 27.2 ± 1.7 at the baseline. Moreover, the mean MMSE score was 21.8 ± 5.2 at 4 years after the baseline and conversion to AD dementia was 59.8% within 4 years of the baseline. There was a significant negative correlation between plasma Aβ 42 /Aβ 40 and plasma P-tau181 (coefficient of determination (R2) = −0.30; P < 0.0001; see Extended Data Fig. 1) and a significant positive correlation between plasma P-tau181 and plasma NfL (R2 = 0.33; P < 0.0001; Extended Data Fig. 1), but no significant correlation was observed between plasma Aβ 42 /Aβ 40 and plasma NfL (R2 = −0.08; P = 0.31; Extended Data Fig. 1). The associations between corresponding CSF biomarkers in the BioFINDER model selection sample are shown in Extended Data Fig. 2. The study procedures are outlined in Fig. 1.

Table 1 Study participant characteristics (model selection) Full size table

Fig. 1: Flow chart of study procedures. In the ADNI cohort, 425 participants had data for plasma P-tau181 and NfL. Of these, 86 also had data for plasma Aβ 42 /Aβ 40 . The number of participants available for different outcomes differed for both BioFINDER and ADNI. Asterisks in the section ’Model selection’ represent the ADNI model selection sample, which included individuals in the ADNI cohort for whom data were available for all plasma biomarkers (that is, Aβ 42 /Aβ 40 , P-tau181 and NfL). In the section ‘Individualized prediction using best models’, ‘internal prognostic validation’ refers to fivefold cross-validation for the internal prognostic validation step, whereas ‘external prognostic validation’ refers to validating the model selection (that is, validating that the best-performing model for BioFINDER was the best-performing model for ADNI, and vice versa). Asterisks in the section ‘Individualized prediction using best models’ represent the ADNI prognostic validation sample, which included individuals in the ADNI cohort for whom data were available for P-tau181 and NfL. Full size image

The data from 86 patients with MCI from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort for whom all three plasma biomarker measurements were available alongside at least one of the primary or secondary outcomes were used to replicate model selection (Table 1). The mean age was 71.5 years, the mean education duration was 16.4 years, 51.2% were female and the mean MMSE score was 28.3 ± 1.7 at the baseline. Moreover, the mean MMSE score was 27.6 ± 2.9 at 4 years after the baseline and conversion to AD dementia was 10.8% within 4 years of the baseline. There was a significant negative correlation between plasma Aβ 42 /Aβ 40 and plasma P-tau181 (R2 = −0.31; P = 0.0023; Extended Data Fig. 2) but no significant correlation between plasma P-tau181 and plasma NfL (R2 = 0.15; P = 0.16; Extended Data Fig. 2) or between plasma Aβ 42 /Aβ 40 and plasma NfL (R2 = −0.16; P = 0.11; Extended Data Fig. 2).

To validate our prognostic models, we used data from 425 patients with MCI from the ADNI cohort for whom plasma P-tau181 and NfL measurements were available alongside at least one of the primary or secondary outcomes available. The mean age of this cohort was 71.0 years, the mean education duration was 16.1 years, 51.8% were female and the mean MMSE score was 28.2 ± 1.7 at the baseline. Moreover, the mean MMSE was 26.6 ± 4.1 4 years after the baseline and conversion to AD dementia was 33.1% within 4 years of the baseline. There was a significant positive correlation between plasma P-tau181 and plasma NfL (R2 = 0.35; P < 0.0001; Extended Data Fig. 3).

With 4-year MMSE as the outcome in the BioFINDER model selection sample (n = 118), the model that included plasma Aβ 42 /Aβ 40 , P-tau181 and NfL as predictors (full model; R2 = 0.36; Akaike information criterion (AIC) = 684) fit the data significantly better than the basic model that included age, sex, education and baseline MMSE score (R2 = 0.24; AIC = 702; P = 0.0001 compared with the full model). However, the best-fitting model according to the AIC was the one that included plasma P-tau181 and NfL but not Aβ 42 /Aβ 40 (R2 = 0.36; AIC = 683). In the best-fitting model, there was a significant individual effect of P-tau181 (β = −1.65; P < 0.0001) but not NfL (β = −0.70; P = 0.13) (Fig. 2a,b and Supplementary Table 1).

Fig. 2: Modeling cognitive decline using plasma Aβ 42 /Aβ 40 , P-tau181 and NfL. Results from modeling cognitive decline in patients with MCI using plasma biomarkers. a, R2 (x axis) and AIC values (numbers in plots) for each plasma-based model, with MMSE evaluated 4 years after the baseline as the outcome, in the ADNI and BioFINDER cohorts. The dashed vertical lines show basic model performances for reference. Error bars indicate 95% confidence intervals. All models also included age, sex, education and baseline MMSE as predictors. b, Coefficients from each plasma-based model, with MMSE evaluated 4 years after the baseline as the outcome, in the ADNI and BioFINDER cohorts. Statistically significant variables, as determined by two-sided t-test on regression coefficients, are plotted with an asterisk instead of a square, and the error bars represent 95% confidence intervals. c, Estimated MMSE trajectories (colored lines), together with the estimated trajectories from the best-fitting model (P-tau181 and NfL; Aβ 42 /Aβ 40 was not taken forward for assessing predictive performance), according to biomarker status, adjusted for age, sex, education and baseline MMSE. Error bars represent 95% confidence intervals on estimated trajectories. No corrections for multiple comparisons were performed. Source data Full size image

With 4-year MMSE as the outcome in the ADNI model selection sample used for replication (n = 64), the full plasma model (R2 = 0.25; AIC = 310) fit the data better than the basic model (R2 = 0.15; AIC = 316; P = 0.01 compared with the full ATN model), and the best-fitting model according to the AIC again included plasma P-tau181 and NfL only (R2 = 0.25; AIC = 309). In the best-fitting model, the individual effect of P-tau181 was nearly significant (β = −0.64; P = 0.06) while the individual effect of NfL was significant (β = −1.02; P = 0.02) (Fig. 2a,b and Supplementary Table 1).

With 4-year conversion to AD as the outcome in the BioFINDER model selection sample (n = 107), the full plasma model that included all three biomarkers (area under the curve (AUC) = 0.88; AIC = 106) fit the data significantly better than the basic model (AUC = 0.70 (0.60–0.79); AIC = 140; P < 0.0001 compared with the full model). The best-fitting model according to the AIC included P-tau181 and NfL but not Aβ 42 /Aβ 40 , resulting in an AUC of 0.88 (0.82–0.95). A maximum specificity of 95.3% could be achieved given at least 50% sensitivity, and a maximum sensitivity of 96.8% could be achieved given at least 50% specificity (Fig. 3a,b and Supplementary Table 2). In the best-fitting model, there was a significant individual effect of P-tau181 (odds ratio (OR) = 5.87; P = 0.0001) but not NfL (OR = 1.73; P = 0.10). Using Cox regression (Supplementary Table 3), the model with P-tau181 and NfL was still the best-fitting model and there was a significant individual effect of both P-tau181 (hazard ratio (HR) = 2.52; P = 0.006) and NfL (HR = 2.70; P = 0.02).

Fig. 3: Modeling clinical conversion using plasma Aβ 42 /Aβ 40 , P-tau181 and NfL. Results from modeling clinical conversion in patients with MCI using plasma biomarkers. a, AUC (x axis) and AIC values (numbers in plots) for each plasma-based model, with conversion to AD within 4 years after baseline as the outcome, in the BioFINDER and ADNI cohorts. The dashed vertical lines show basic model performances for reference. Error bars indicate 95% confidence intervals. All models also included age, sex, education and baseline MMSE as predictors. b, Coefficients from each plasma-based model, with conversion to AD within 4 years after the baseline as the outcome, in the BioFINDER and ADNI cohorts. Statistically significant variables, as determined by two-sided t-test on regression coefficients, are plotted with an asterisk instead of a square, and the error bars represent 95% confidence intervals. c, Estimated probability of not converting to AD, as predicted from the best-fitting model (P-tau181 and NfL; Aβ 42 /Aβ 40 was not taken forward for assessing predictive performance), according to biomarker status, adjusted for age, sex, education and baseline MMSE. Error bars represent 95% confidence intervals on estimated trajectories. Source data Full size image

With 4-year conversion to AD as the outcome in the ADNI model selection sample used for replication (n = 74), the full plasma model (AUC = 0.88 (0.80–0.98); AIC = 50) fit the data better than the basic model (AUC = 0.74 (0.52–0.95); AIC = 57; P = 0.005 compared with the full ATN model) and the best-fitting model according to AIC again included P-tau181 and NfL only. The AUC for this model was 0.89. A maximum specificity of 90.1% could be achieved given at least 50% sensitivity, and a maximum sensitivity of 100% could be achieved given at least 50% specificity (Fig. 3a,b and Supplementary Table 2). In the best-fitting model, the individual effect of P-tau181 was significant (OR = 4.58; P = 0.009) while the individual effect of NfL was not significant (OR = 2.15; P = 0.20). Using Cox regression (Supplementary Table 3), the model with P-tau181 and NfL was again still the best-fitting model, with a significant effect in this model for P-tau181 (HR = 2.22; P < 0.0001) but not NfL (HR = 1.27; P = 0.13). Receiver operating characteristic (ROC) curves for the main models are presented in Fig. 4. ROC curves for all of the models in the model selection analysis with 4-year conversion to AD dementia as the outcome are presented in Extended Data Fig. 4.

Fig. 4: ROC curves for clinical conversion across cohorts. a,b, ROC curves from modeling clinical conversion in patients with MCI using plasma biomarkers in the BioFINDER (a) and ADNI (b) cohorts. The basic model (age, sex, education and baseline MMSE) is shown along with the best-performing model selected for further prognostic validation (basic model + plasma P-tau181 + plasma NfL (plasma TN)). Results from the full CSF model (basic model + CSF Aβ 42 /Aβ 40 + CSF P-tau181 + CSF NfL (CSF ATN)) are also shown for the BioFINDER cohort. These results are not shown for the ADNI cohort as the full CSF model was not available (but note that the CSF TN model for ADNI was tested at the prognostic validation stage). Full size image

Effect of the APOE ε4 genotype on model selection

Including APOE ε4 genotype status (that is, ε4 carrier versus non-carrier) in the basic model did not substantially affect model selection for the primary cognitive outcome (Supplementary Table 4) or for the primary clinical outcome (Supplementary Table 5). Moreover, the full plasma ATN model still outperformed the basic model for both the primary cognitive and clinical outcomes.

Sensitivity analysis using secondary and exploratory outcomes

We performed the same model selection procedure as outlined above using the secondary and exploratory outcomes. We found that the best-fitting models identified here varied across outcomes, whereas the full plasma model including Aβ 42 /Aβ 40 , P-tau181 and NfL still always outperformed the basic model alone (Supplementary Tables 4–13).

Sensitivity analysis using an alternative plasma Aβ 42 /Aβ 40 assay

Because plasma Aβ 42 /Aβ 40 was not selected as a predictor in the best-fitting models for the co-primary outcomes, we tested whether this result differed when using a mass spectrometry assay (Araclon Biotech) instead of the Elecsys assay used in the BioFINDER cohort. Here, the best-fitting models according to AIC when using the mass spectrometry plasma Aβ 42 /Aβ 40 assay still did not include Aβ 42 /Aβ 40 . Instead, the best-fitting model with 4-year MMSE as the outcome included P-tau181 and NfL only and the best-fitting model with 4-year conversion to AD as the outcome included P-tau181 only (Supplementary Tables 12 and 13).

Sensitivity analysis using plasma P-tau217 in the BioFINDER cohort

Primary analyses (that is, 4-year MMSE and 4-year conversion to AD) were repeated using plasma P-tau217 (see Palmqvist et al.18 for an assay description) instead of P-tau181. As plasma P-tau217 data are not available for ADNI, these analyses were performed using BioFINDER participants only. Findings from these analyses replicated those found when using P-tau181 (Supplementary Tables 14 and 15): P-tau in combination with NfL remained the best model for both outcomes, and R2 and AUC values were similar when using P-tau217 or P-tau181 (for MMSE, compare Supplementary Table 14 with Supplementary Table 1; for conversion to AD, compare Supplementary Table 15 with Supplementary Table 2).

Study population characteristics (prognostic validation sample)

Since the model that included both plasma P-tau181 and NfL (but not Aβ 42 /Aβ 40 ) provided the best fit across co-primary outcomes in both cohorts, this model was taken forward in the prognostic validation stage. For this analysis, we therefore only required participants to have available data on plasma P-tau181 and NfL measurements and at least one of the primary or secondary outcomes. As such, 148 patients with MCI from the BioFINDER cohort (no difference from the BioFINDER model selection sample) and 425 patients with MCI from the ADNI cohort (Table 2; see Extended Data Fig. 5 for the association between plasma P-tau181 and NfL in this group) were included.

Table 2 Study participant characteristics in ADNI (prognostic validation) Full size table

We performed an internal validation where the patient-level predictive performance of the best-fitting plasma model (the basic model plus plasma P-tau181 and NfL) was evaluated within each cohort and compared with the basic model alone and with the full CSF model (the basic model plus CSF Aβ 42 /Aβ 40 , P-tau181 and NfL).

With 4-year MMSE as the outcome in the BioFINDER prognostic validation sample (n = 118), the best-fitting plasma model improved cross-validated, out-of-sample prediction compared with the basic model alone (mean absolute error (MAE) = 3.07 points versus 3.36 points; P < 0.001; 8.5% improvement) and showed no significant difference compared with the full CSF model (P = 0.68 over 1,000 bootstrapped trials). With 4-year MMSE as the outcome within the ADNI prognostic validation sample (n = 252), the same best-fitting plasma model improved out-of-sample prediction compared with the basic model alone (MAE = 2.42 points versus 2.49 points; P < 0.001; 2.9% improvement).

With 4-year conversion to AD as the outcome in the BioFINDER prognostic validation sample (n = 107), the best-fitting plasma model improved out-of-sample prediction compared with the basic model alone (AUC = 0.63 versus 0.83 (note that these cross-validated AUCs are, as expected, lower than the AUCs from corresponding models fit on all data for model selection); P < 0.001; 31.7% improvement) and significantly outperformed the full CSF model (P = 0.002 over 1,000 bootstrapped trials; 5% improvement). With 4-year conversion to AD as the outcome in the ADNI prognostic validation sample (n = 320), the plasma model significantly improved out-of-sample prediction compared with the basic model alone (AUC = 0.66 versus 0.76; P < 0.001; 15.4% improvement). For a visual depiction of individualized predicted probabilities of conversion to AD across models, see Fig. 5.

Fig. 5: Individualized prediction of 4-year conversion from MCI to AD dementia. a,b, Results from internal cross-validation for clinical conversion for the best-performing models, as identified in the first stage of analysis using all available BioFINDER (a; n = 107) and ADNI (b; n = 320) patients. The box plots show the minimum and maximum values (whiskers), as well as the median (central line) and interquartile range (that is, the first quartile (25th percentile; bottom edge of box) and third quartile (75th percentile; top edge of box)). The values plotted here show the predicted probability of conversion from MCI to AD dementia for each individual in the BioFINDER and ADNI cohorts, showing 32.3% improvement (AUC = 0.62 for the basic model versus AUC = 0.82 for the P-tau181 and NfL model) of the plasma-based model over the basic model in BioFINDER and a 15.4% improvement (AUC = 0.66 for the basic model versus AUC = 0.76 for the P-tau181 and NfL model) in ADNI. Full size image

We performed an external validation whereby the patient-level predictive performance of the best-fitting plasma model (the basic model plus plasma P-tau181 and NfL) was evaluated across each cohort by first fitting the model on BioFINDER participants and then testing on ADNI participants, and vice versa. For this analysis, biomarkers were dichotomized.

With 4-year MMSE as the outcome in the BioFINDER (n = 118, of whom 28 were T−N−, 13 were T−N+, 46 were T+N− and 31 were T+N+) and ADNI prognostic validation samples (n = 243, of whom 118 were T−N−, 35 were T−N+, 46 were T+N− and 44 were T+N+), the plasma model significantly improved prediction on the test cohort compared with the basic model, both when the model was fit on BioFINDER and tested on ADNI (MAE = 3.74 versus 4.08; P = 0.0006; 8.3% improvement) and when the model was fit on ADNI and tested on BioFINDER (MAE = 4.15 versus 5.19; P < 0.0001; 20.1% improvement).

With 4-year conversion to AD as the outcome in the BioFINDER (n = 107; of whom 20 were T−N−, five were T−N+, 49 were T+N− and 33 were T+N+) and ADNI prognostic validation samples (n = 314, of whom 139 were T−N−, 45 were T−N+, 62 were T+N− and 68 were T+N+), the plasma model improved prediction on the unseen cohort both when the model was fit on ADNI and tested on BioFINDER (AUC = 0.61 versus 0.79; P < 0.0001; 29.3% improvement) and when the model was fit on BioFINDER and tested on ADNI (AUC = 0.62 versus 0.73; P < 0.0001; 18.3% improvement).