Error unawareness as a driver of cognitive decline in individuals at risk for Alzheimer’s disease

Abstract

Anosognosia, or unawareness of one’s deficits, predicts faster clinical progression in Alzheimer’s disease (AD), though its mechanisms remain unclear. We investigated five-year trajectories of response awareness and error-monitoring efficiency using second-order judgments from an annual word recognition memory task in cognitively unimpaired individuals at risk for AD from the INSIGHT-preAD cohort. Participants were divided into three groups according to amyloid status at baseline (positive, A + ; negative, A–) and clinical progression (controls, CTRL A+ or A–; and progressors, PROG). Task accuracy declined in PROG, while CTRL A+ showed continued learning, and CTRL A– remained accurate. Inefficient error-monitoring predicted subsequent performance decline in the PROG, with a strong interaction between accuracy and error-monitoring already at baseline. This effect was primarily driven by error unawareness. These findings provide first evidence that error unawareness — a core feature of anosognosia — may foster cognitive decline through the disruption of error-feedback learning mechanisms, potentially accelerating AD progression.

This study advances understanding of AD progression by linking anosognosia-related error unawareness to disrupted error-feedback learning and cognitive decline from the preclinical stage, with implications for disease modeling, prognosis, and therapy.

Introduction

Learning is a dynamic process that depends on the brain’s capacity to monitor outcome, evaluate performance, and adjust behavior accordingly. Errors play a central role in this adaptive mechanism by providing critical feedback signals that guide learning. When an error occurs, the brain generates rapid electrophysiological responses, such as the error-related negativity (ERN) and the error positivity (Pe), which reflect the detection and the awareness of errors, respectively^1–4. The anterior and posterior cingulate cortices (ACC, PCC) constitute key nodes of this error-monitoring network², supporting the integration of feedback and the updating of cognitive models underlying learning⁵. Awareness of errors enhances the encoding of corrective feedback, leading to improved performance and memory retention⁶. Conversely, unawareness of errors disrupts these mechanisms, resulting in suboptimal behavioral adjustment and reduced cognitive flexibility.

Typically, patients with Alzheimer’s disease (AD) show a reduced ability to recognize their own errors, reflecting an early manifestation of a broader loss of awareness into their cognitive and functional difficulties—a phenomenon known as anosognosia^7–9. Consistent with this, deficits in error-monitoring have been observed from the earliest AD stages in the same longitudinal cohort on which this study is based¹⁰. In particular, we have found that neural markers of error awareness—specifically, the Pe amplitude—decreased in individuals who subsequently developed AD within the INSIGHT-preAD cohort. Notably, these individuals presented significantly lower Pe amplitudes at the time of diagnosis compared to study entry, which occurred alongside a decline in their abilities to monitor their cognitive difficulties, as assessed by an offline measure of anosognosia¹⁰. Although the limited sample size precluded formal correlation analyses between these two variables, preventing any inference of a direct relationship, the temporal co-occurrence of these changes is nevertheless consistent with evidence linking Pe amplitude to awareness-related processes^2,4. This pattern distinguished these individuals from those who remained cognitively stable over the study period.

Similarly, learning impairments emerge very early in the AD continuum¹¹, and some evidence suggest that they can even precede hippocampal-dependent memory dysfunction. For example, learning-slope measures have demonstrated that subtle reductions in the acquisition rate across repeated trials can differentiate amyloid-positive from amyloid-negative individuals, even when they are still cognitively unimpaired^12,13.

The current study integrates these two lines of evidence. We hypothesized that the inability to recognize errors (for example, while performing a cognitive task) could disrupt the mechanism by which errors are used to inform learning and adjust behavior, potentially leading to cognitive decline from the earliest stages of AD¹⁴. Elucidating this mechanism is expected to provide systems-level insight into the factors that underlie heterogeneity in progression dynamics along the Alzheimer’s continuum.

To address this critical issue, we investigated longitudinal changes in self-monitoring in elderly individuals at risk for AD from the INSIGHT-preAD cohort¹⁵. Participants were divided into three groups according to amyloid status at baseline (positive, A+; negative, A–) and clinical progression (controls, CTRL A+ or A–; and progressors, PROG), with full details presented in the Methods section. Specifically, we examined five-year trajectories of response awareness and error-monitoring, using second-order judgments during an annual word recognition memory task of 208 trials per visit.

At each study visit, participants were asked whether they had previously seen the word presented on the screen, followed after each trial by second-order judgments of response certainty^16,17. These judgments were categorized as Correct Sure (CS), Correct Not Sure (CNS), Error Sure (ES; indicative of error unawareness), or Error Not Sure (ENS). Longitudinal changes in word recognition performance were quantified using the Task Performance Index (TPI), defined as the difference in task accuracy between the last and first visits. Note that the first visit corresponds to month 0 (i.e., M0) to the three groups, while the last visit corresponds to month 60 (i.e., Mlast) for the control groups and to the visit at which the AD diagnosis was established (i.e., Mdiag) for the PROG group. To examine the impact of self-monitoring on task performance over time, we computed the Self-Monitoring inefficiency Index (SMI), derived from two key indicators of impaired metacognitive monitoring (ES and CNS)¹⁸. Specifically, the SMI reflects the combined frequency of ES and CNS events, normalized by the total number of trials (for details, see the Methods section). Our findings indicate that error unawareness may mechanistically link inefficient error-monitoring to early learning difficulties in preclinical AD, which can potentially accelerate cognitive decline by disrupting error-feedback signals and limiting the implementation of compensatory strategies to counter emerging deficits. This interpretation of the findings should be considered in light of evidence that memory processes and memory self-monitoring rely on partially dissociable systems that can be differentially impaired¹⁹. In brief, this study advances a novel mechanistic account that may enhance our understanding of clinical progression along the Alzheimer’s continuum.

Results

Clinical and demographic characteristics of the three groups

The PROG group comprises the 15 amyloid-positive individuals who progressed to prodromal AD throughout the five-year study period, thereby constituting our primary group of interest. The two control groups were randomly selected from individuals who remained cognitively stable despite being either amyloid-positive (CTRL A+) or amyloid-negative (CTRL A–). As shown in Table 1, the three groups were matched in age, sex, and education level. In the PROG group, 9 of 15 individuals (60%) were APOE ɛ4 carriers, compared with 6 of 15 (40%) in the CTRL A+, and 4 of 15 (27%) in the CTRL A– group. Despite these numerical differences, the groups did not differ significantly in APOE ε4 frequency (p = 0.18).

Table 1.

Clinical and demographic characteristics of the three groups

	All	CTRL A–	CTRL A+	PROG	P value
	n = 45	n = 15	n = 15	n = 15
I. Baseline (M0)
Age (years)	77.5 (2.9)	77.9 (2.4)	76.6 (2.6)	78.0 (3.5)	0.092
Sex, F/M (%F)	26/19 (57.8%)	9/6 (60%)	9/6 (60%)	8/7 (53.3%)	0.91
Years of education	12.7 (5.3)	14.1 (5.4)	10.7 (5.1)	13.3 (5.1)	0.098
MMSE score	28.5 (0.8)	29 (0.7)c	28.3 (0.9)	28.2 (0.7)a	0.015
FCSRT – Total recall score	46.1 (2.1)	47.7 (0.5)bc	45.9 (2.0)a	44.7 (2.2)a	<0.001
DMS-48 – 1-hour delayed score	46.0 (2.0)	46.7 (1.6)	45.5 (2.3)	45.9 (2.1)	0.31
FAB	16.2 (2.0)	16.9 (2.0)	15.7 (1.9)	16.1 (2.1)	0.11
TMT B-A (seconds)	57.4 (45.1)	35.6 (24.2)	73.3 (59.8)	63.4 (37.4)	0.034
GDS - 15 items	2.1 (2.5)	2.1 (2.6)	1.9 (2.4)	2.4 (2.7)	0.91
STAI-Y-B	39.4 (10.4)	39.7 (10.8)	38.9 (9.7)	39.5 (11.4)	0.97
SAS	8.7 (4.0)	9.1 (5.9)	7.5 (3.0)	9.5 (2.1)	0.16
APOE genotype					0.45
E2/E3	3 (6.7%)	2 (13.3%)	1 (6.7%)	0 (0%)
E2/E4	1 (2.2%)	0 (0%)	1 (6.7%)	0 (0%)
E3/E3	23 (51.1%)	9 (60%)	8 (53.3%)	6 (40%)
E3/E4	15 (33.3%)	4 (26.7%)	4 (26.7%)	7 (46.7%)
E4/E4	3 (6.7%)	0 (0%)	1 (6.7%)	2 (13.3%)
APOE allele ɛ4, Presence/Absence (%Presence)	19/26 (42.2%)	4/11 (26.7%)	6/9 (40%)	9/6 (60%)	0.18
II. Mlast/Mdiag
Age (years)	82.0 (2.8)	82.3 (2.6)	81.6 (2.6)	82.1 (3.2)	0.38
MMSE score	27.7 (2.8)	29.1 (0.7)c	28.1 (1.9)	25.9 (3.9)a	0.0041**
FCSRT – Total recall score	41.3 (8.2)	47.8 (0.4)bc	43.6 (1.6)ac	32.5 (8.7)ab	<0.001***
DMS-48 – 1-hour delayed score	45.7 (2.6)	47.3 (0.8)c	46.2 (1.4)c	43.5 (3.3)ab	<0.001***
FAB	15.6 (1.9)	17.1 (1.3)bc	15.6 (1.7)ac	14.2 (1.6)ab	<0.001***
TMT B-A (seconds)	62.1 (38.6)	43 (25.5)	72.2 (45.4)	71.8 (34.5)	0.037*
GDS-15 items	2.5 (2.6)	2.1 (2.2)	2.5 (2.4)	2.7 (3.3)	0.82
STAI-Y-B	38.5 (10.3)	39 (10.8)	38.7 (9.9)	37.6 (10.7)	0.91
SAS	10.8 (3.5)	11 (5.0)	10.7 (2.7)	10.6 (2.5)	0.96
III. Δ (Mlast/Mdiag – M0)
MMSE score	−0.8 (2.7)	0.1 (0.8)c	−0.2 (1.8)	−2.3 (3.8)**,a	0.039*
FCSRT – Total recall score	−4.8 (7.8)	0.1 (0.4)bc	−2.3 (2.3)**,ac	−12.2 (9.6)***,ab	<0.001***
DMS-48 – 1-hour delayed score	−0.4 (2.9)	0.6 (0.8)c	0.7 (1.4)	−2.4 (3.3)*,a	0.026*
FAB	−0.6 (2.6)	0.2 (2.7)	−0.1 (2.3)	−1.9 (2.6)*	0.084
TMT B-A (seconds)	5.2 (36.7)	7.4 (28.4)	−1.1 (45.4)	9.6 (34.5)	0.62

MMSE Mini-Mental State Examination²¹, FCRST Free and Cued Selective Reminding Test²⁰, DMS-48 Delayed Matching to Sample-48 items, FAB Frontal Assessment Battery²², TMT Trail Making Test²⁷, GDS-15 items Geriatric Depression Scale²³, STAI-Y-B State-Trait Anxiety Inventory (B, trait scale)^24,25, SAS Starkstein Apathy Scale²⁶.

Continuous variables are presented as mean (standard deviation). P values indicate between-group comparisons using Fisher’s exact test for Sex and the Kruskal–Wallis test for continuous variables, followed by Dunn’s post hoc test with Holm adjustment.

Superscripts a, b, and c indicate significant differences from groups CTRL A–, CTRL A+, and PROG, respectively.

Within-group changes (Δ values) between M0 and Mlast/Mdiag were evaluated using the Wilcoxon signed-rank test and are reported as *p < 0.05, **p < 0.01, and ***p < 0.001 where significant.

At baseline (M0), all participants were cognitively unimpaired. Interestingly, the PROG and the CTRL A+ groups showed lower performances regarding total recall scores in the Free and Cued Selective Reminding Test (FCSRT)²⁰ relative to the CTRL A– (p < 0.001). Still, all three groups performed within the normal range, indicating accurate episodic memory functioning. The PROG group exhibited slightly reduced global cognition relative to CTRL A–, as measured by the Mini-Mental State Examination (MMSE)²¹ (p = 0.015).

At Mlast/Mdiag, between-group differences were accentuated (MMSE: p = 0.0041; FCSRT: p < 0.001), with the PROG group showing pathological scores (MMSE: 25.9 ± 3.9; FCSRT: 32.5 ± 8.7; mean ± standard deviation [SD]) and within-group declines (MMSE: –2.3 ± 3.8, p = 0.0098; FCSRT: –12.2 ± 9.6, p < 0.001), while the CTRL A– had no change, and the CTRL A+ presented only modest FCSRT decline (–2.3 ± 2.3, p = 0.0067). Additionally, the PROG and the CTRL A+ groups showed the emergence of mild dysexecutive difficulties, as indicated by mean Frontal Assessment Battery (FAB) scores²² of 14.2 and 15.6, respectively. In contrast, the CTRL A– group exhibited a normal mean score of 17.1. Between-group differences were significant (p < 0.001).

No neuropsychiatric symptoms, including depression, anxiety or apathy, were observed in any of the three groups, as assessed using the Geriatric Depression Scale²³, the State-Trait Anxiety Inventory^24,25 and the Starkstein Apathy Scale²⁶.

Overall, a decline in episodic memory, as assessed by the FCSRT, clearly distinguished the PROG group from both the CTRL A+ and the CTRL A– groups at Mlast/Mdiag compared to baseline (M0). Decline in global cognitive status, as measured by the MMSE, was less discriminative. In contrast, executive functions, as assessed by the FAB or the Trail Making Test (TMT)²⁷, did not differ significantly from M0 to Mlast/Mdiag in between-group comparisons [Table 1, subsection: III. Δ (Mlast/Mdiag – M0)].

Second-order judgment for the word recognition memory task: four response types

Figure 1 shows the certainty judgment patterns (“Sure” vs. “Not sure”) regarding either erroneous or correct responses in the annual word recognition memory task across visits for the three groups within the five-year study period. The PROG group committed more errors than the CTRL groups over time, with errors predominantly associated with increased unawareness (i.e., increased number of erroneous responses for which the individuals were “Sure”, ES, that is: error unawareness). This was accompanied by a decline in task accuracy over time, reflected by fewer correct responses, which, in contrast to erroneous responses, were associated with increased uncertainty (i.e., increased number of correct responses for which the individuals were “Not sure”, CNS) and decreased awareness (i.e., decreased number of correct responses for which the individuals were “Sure”, CS). Of note, in the PROG group, decline in performance on this word recognition memory task (TPI, task performance index) was correlated with decline in the FCSRT total recall from M0 to Mdiag (ρ = 0.55, p = 0.033).

Fig. 1. Certainty response patterns (“Sure” vs. “Not sure”) for erroneous and correct responses in the annual word recognition memory task across study visits in the three groups (n = 15 participants per group).

a Predicted trajectories of erroneous and b correct “Sure” vs. “Not sure” responses over five years in the CTRL A–, CTRL A + , and PROG groups. Curves were derived from Poisson generalized linear mixed models (GLMMs); solid lines denote “Sure” responses, dashed lines denote “Not sure” responses, and shaded areas indicate 95% confidence intervals around trajectories. Legends report the estimated annual percentage change (± 95% CI) for each response type and group, with significance levels indicated (**p < 0.01, ***p < 0.001). Post hoc simple linear slope comparisons using emmeans, with p values adjusted using the Tukey method for multiple comparisons, showed that PROG differed from both CTRL groups across error and correct certainty patterns, while CTRL A– and CTRL A+ generally did not differ, except for Correct “Not sure”. Only PROG showed significant differences between “Sure” and “Not sure” slopes. Details of the slope estimations and comparisons are provided in Supplementary Tables 1(ii)–(iii) and 2(ii)–(iii).

The CTRL A+ group improved their performance over time, with correct responses increasingly classified as “Sure” (i.e., a higher number of correct responses for which individuals were aware, CS) rather than “Not sure”, while error unawareness (ES) tended to disappear. The CTRL A– group maintained high accuracy (near 100%) throughout the study period, accompanied by a predominantly correct “Sure” response pattern.

In summary, both ES, relative to ENS, and CNS, relative to CS, increased significantly in the PROG group over time, whereas the inverse pattern was observed in the CTRL A+ group. The CTRL A– group showed a stable pattern across time. This is particularly relevant because both ES and CNS reflect a miscalibrated error-monitoring system.

Longitudinal modulation of certainty judgments for erroneous and correct responses

For erroneous responses, the Poisson generalized linear mixed models (GLMM) revealed significant main effects of Certainty, Time of visits (continuous), and Group, as well as interactions, including the three-way interaction (Certainty × Time × Group: χ²(2) = 20.7, p < 0.001), indicating that the modulation of errors by response certainty varied across groups and over time. At M0, the Sure–Not sure difference (ES − ENS) was largest in PROG (10.88 ± 1.91), intermediate in CTRL A+ (5.67 ± 1.07), and smallest in CTRL A– (1.22 ± 0.36), all p < 0.001. By Mlast/Mdiag, the difference increased substantially in PROG (28.07 ± 5.09), while CTRL groups showed only modest changes (CTRL A+: 4.09 ± 0.79; CTRL A–: 1.56 ± 0.41). Diff-in-diff comparisons confirmed that PROG differed significantly from the two CTRL groups at both time points. Over time, the change in the Sure–Not sure difference was significant only in PROG (17.2 ± 3.64, p < 0.001), indicating a pronounced temporal increase between M0 and Mdiag in the Sure–Not sure error disparity in this group. Detailed results of the tests based on GLMMs are presented in Supplementary Tables 1(i)–(iv).

For correct responses, the three-way interaction was also significant (Certainty × Time × Group: χ²(2) = 269.7, p < 0.001), with diff-in-diff analyses indicating that reductions in the Sure–Not sure delta following correct responses over time were significantly robust in PROG (−53.17 ± 4.72, p < 0.001). In contrast, CTRL A+ showed an inverted pattern with moderate increase in the Sure–Not sure delta following correct responses (+23.94 ± 4.42, p < 0.001), and CTRL A− remained stable (−0.54 ± 4.97, p = 0.91). See Supplementary Tables 2(i)–(iv) for full GLMM results.

Link between accuracy and self-monitoring inefficiency index with component contributions

The SMI was strongly associated with task accuracy (χ²(1) = 146.7, p < 0.001). This relationship varied by group (SMI × Group: χ²(2) = 48.4, p < 0.001), with modest changes across visits (categorical, M0 to M60; SMI × Visit: χ²(5) = 49.6, p < 0.001), and post-hoc slopes indicating that the SMI–Accuracy association remained largely stable across visits. The three-way interaction was not significant (SMI × Group × Visit: χ²(10) = 15.1, p = 0.13). Trend analyses indicated that SMI was negatively associated with accuracy already at baseline, and remained consistent across visits in the PROG group. This trend was not observed in the two CTRL groups. Pairwise comparisons confirmed that the negative SMI–Accuracy relationship was significantly stronger in PROG than in CTRL A+ at most time points, with significant differences from M0 to M48 and a non-significant trend at M60 (p = 0.094). Test details are provided in Supplementary Table 3(i)–(iii).

A sensitivity analysis excluding three CTRL A+ patients exhibiting outlying values in their change in SMI between baseline (M0) and Mlast (M60) slightly altered the results for this group. In this analysis, the three-way interaction became significant (SMI × Group × Visit: χ²(10) = 21.6, p = 0.017), and the SMI–Accuracy association varied across visits, being significantly negative at M0, M24, M36, and M60, but not at M12 and M48, at which time points the difference with the PROG group was maintained. These results likely reflect heterogeneity among CTRL A+ participants, most of whom maintained their self-monitoring abilities over time, while a few even showed an improvement over time. Details of the sensitivity analysis are provided in Supplementary Table 4(i)–(iii).

In the PROG group, decomposition of SMI into its components (ES and CNS) revealed that ES was the dominant predictor of accuracy. Adding ES to a Visit-only model strongly improved fit (χ²(1) = 207.9, p < 0.001), whereas adding CNS alone did not (χ²(1) = 0.04, p = 0.85). When both components were included, CNS provided a small but still significant additional effect beyond ES (χ²(1) = 30.8, p < 0.001), and ES remained the primary driver (χ²(1) = 238.7, p < 0.001). No significant interactions with Visit were observed for either component. Relative-contribution analyses confirmed this asymmetry: based on log-likelihood differences, ES accounted for 94.5% of the combined contribution vs. 5.5% for CNS; using marginal R², ES explained 98.1% of the variance attributable to SMI components, with CNS contributing only 1.9%. Model comparisons and detailed statistics for all nested models, including contributions of ES and CNS and interactions with Visit, are reported in Supplementary Table 5.

Correlation between task performance index and changes in self-monitoring inefficiency index: last visit vs. baseline

As shown in Fig. 2, changes in SMI (Mlast/Mdiag – M0) and TPI were strongly negatively correlated in the PROG group (ρ = −0.91, p < 0.001), indicating that self-monitoring inefficiency (positive SMI changes) played a critical role in task performance decline (negative TPI changes) from baseline to Mdiag. A negative correlation was also found in CTRL A– (ρ = −0.91, p < 0.001), but the very limited variability within this group appears to primarily reflect the relative homogeneity of the CTRL A– group with respect to changes in SMI and TPI, as illustrated by the corresponding box plots in Fig. 2. In contrast, no correlation was observed in the CTRL A+ group (ρ = −0.12, p = 0.67), likely reflecting greater heterogeneity in changes in SMI compared with the CTRL A– group, as further illustrated by the box plots in Fig. 2.

Fig. 2. Changes in self-monitoring inefficiency index and their influence on task accuracy across groups between M0 and Mlast/Mdiag (n = 15 participants per group).

Spearman’s correlation coefficient (ρ) and corresponding significance (p) indicate the association between task performance index (TPI) and Self-Monitoring inefficiency Index (SMI) for each group. Individual points represent participants’ SMI and TPI values. Marginal distributions of SMI and TPI by group are shown as box plots along the top and right sides of the figure; the center line indicates the median, box limits represent the first and third quartiles, whiskers extend to the most extreme values within 1.5× the interquartile range, and outliers are marked by “×” symbols. A sensitivity analysis excluding these outliers further confirmed the correlations (CTRL A–: ρ = −0.88, p < 0.001; CTRL A + : ρ = −0.31, p = 0.32; PROG: ρ = −0.88, p < 0.001).

Discussion

Understanding the early clinical progression over the Alzheimer’s continuum represents a major public health challenge in the context of an aging global population. In this study, we investigated five-year trajectories of response awareness and error-monitoring in older individuals at risk for AD, using second-order judgments collected annually during a word recognition memory task within a five-year period.

Recognition memory relies primarily on medial temporal structures, namely the hippocampus and the adjacent perirhinal cortex²⁸, whereas real-time performance monitoring engages fronto-cingulate networks^5,19. This is reflected in electrophysiological markers of error awareness, such as the Pe amplitude, which further highlights the neural dissociation between these processes¹⁰. The word recognition memory task used in this study draws on both processes. By providing a direct behavioral index of real-time memory self-monitoring, the second question of the task was expected to reveal disruptions in self-evaluative processes that contribute to the emergence of anosognosia, even before memory deficits become pronounced.

Our results show that self-monitoring inefficiency is tightly linked to the degree of task performance decline in individuals at risk who later progressed to AD. Notably, in the PROG group, poorer self-monitoring efficiency predicted a steeper decline over time, and a strong interaction between accuracy and self-monitoring was already present at baseline. This interaction was primarily explained by frequent error unawareness, though uncertainty after correct responses also contributed. Thus, individuals who progressed to AD along the study period exhibited detectable self-monitoring deficits from the outset, despite being cognitively unimpaired at this time point. These inefficiencies differentiated the PROG group more clearly than time-dependent changes alone.

It is, however, noteworthy that although most CTRL A+ participants maintained stable SMI, three showed an improvement in their self-monitoring abilities, introducing heterogeneity that influenced group-level effects. Excluding these outliers revealed a stronger, time-varying SMI–Accuracy relationship and a significant three-way interaction (SMI × Group × Visit). These findings seem of particular interest because they suggest that among these amyloid-positive controls, self-monitoring heterogeneity may be associated with distinct long-term (beyond this five-year study period) cognitive trajectories. This resembles variability in individuals’ brain resilience across the long AD continuum²⁹.

The PROG group did not benefit from annual task repetition, unlike the CTRL A+ group, who showed continued learning suggesting a practice effect. This is consistent with previous evidence indicating that preclinical AD is marked by a reduction in practice effects in episodic memory³⁰. Individuals of CTRL A– group remained consistently accurate throughout the study.

These findings seem to indicate that inefficient self-monitoring may be a trigger of cognitive decline in individuals at risk for AD, with error unawareness playing a key role in delineating Alzheimer’s trajectories through the disruption of error-feedback learning processes. This mechanism could perpetuate a vicious circle of cognitive decline, by preventing individuals from implementing compensatory strategies to counteract emerging difficulties, potentially driving clinical disease progression¹⁴.

The dual-path model of anosognosia³¹ provides a valuable framework for understanding these findings. It identifies error unawareness as a core feature of anosognosia and proposes two complementary pathways for the emergence of this intriguing syndrome: a direct-path, involving a synaptic failure within the error-monitoring system, particularly affecting the PCC, and an indirect-path, in which disrupted connectivity between emotional–evaluative regions (e.g., amygdala, orbitofrontal cortex) and the error-monitoring related regions impairs the accurate evaluation of errors.

Therefore, unlike tau pathology, which more directly tracks cognitive deficits³², early amyloid accumulation in the PCC may particularly contribute to AD progression^33,34 by disrupting a major hub of the error-monitoring system and the default mode network (DMN)³⁵, leading to cumulative learning difficulties that accelerate cognitive decline and limit the implementation of compensatory strategies¹⁴. Further supporting this view, two key lines of evidence have recently emerged: first, amyloid burden has been associated with gradual decline in self-awareness of memory deficits beginning about three years before Alzheimer’s dementia³⁶; second, functional connectivity impairment within the DMN predicts longitudinal memory decline, synergizing with regional amyloid burden³⁷.

Neuropsychiatric symptoms, including depression, anxiety and apathy, among others, are known to occur along the Alzheimer’s continuum³⁸, with evidence showing that they can influence performance monitoring by affecting either detection and awareness of errors^39,40. This is in line with the indirect-path of our anosognosia model³¹. Within the limits of our sample size, the individuals of the PROG group showed no significant disturbances regarding psychological or behavioral characteristics relative to either control group at baseline, as well as at the time of diagnosis (Table 1). Similarly, executive-attentional functioning is expected to influence performance monitoring. Although both PROG and CTRL A+ groups exhibited slightly lower FAB scores compared to CTRL A– at Mlast/Mdiag, the three groups did not differ significantly in how their executive-attentional performance changed from M0 to Mlast/Mdiag (Table 1). This may be partially explained by the high educational level that characterizes the INSIGHT-preAD cohort, with no significant differences in the number of years of formal education between the three groups^41,42. In turn, this could have an impact on both structural and functional brain networks allowing some individuals to resist or compensate for neurodegeneration^43,44. Together, these observations render such potential confounds unlikely to be the primary contributors of the self-monitoring inefficiency observed in the PROG group, while acknowledging that they could contribute to some extent.

Another relevant issue concerns the apolipoprotein E (APOE) genotype. It is well-established that the APOE ε4 allele increases the risk of AD and is associated with an earlier age of onset by 7–15 years, with homozygous carriers at the highest risk⁴⁵. In our study, the frequency of the APOE ε4 allele did not differ significantly between groups (p = 0.18), indicating comparable genetic risk in this sample. Nevertheless, the distribution of APOE genotypes (Table 1) may contribute to inter-individual variability in brain resistance and resilience to Alzheimer’s pathology^46,47.

Last, our results open a window of opportunity for implementing neurorehabilitation strategies to enhance error-monitoring⁴⁸ and promote adaptive compensatory mechanisms, alongside neuromodulation interventions^49–51 to support healthy brain aging and ultimately slow disease progression.

The relatively small subgroup sizes, resulting from the limited number of individuals who progressed to AD and from criteria specifically related to the study’s hypothesis, constitute the main limitation of this work. This is, however, partially mitigated by the depth and quality of the dataset, derived from a well-characterized monocentric longitudinal cohort with repeated visits. Another limitation is the high education level of the INSIGHT-preAD cohort, as it does not reflect the educational distribution of the general population. Furthermore, we recognize that the lack of correlation analyses between error unawareness (ES factor), used here as a proxy for online anosognosia, and an independent measure of offline anosognosia limits the extent to which the present findings can be interpreted.

In conclusion, this study provides seminal evidence that error unawareness may link self-monitoring inefficiency to the emergence of learning difficulties, potentially driving cognitive decline in individuals at risk for AD. This novel mechanistic account refines current models of Alzheimer’s pathophysiology, helping to explain why anosognosia may predict faster clinical progression. Further confirmation in larger samples is required to fully validate the proposed mechanism. Finally, this study offers a foundation for early interventions targeting the self-monitoring system with the potential to delay the course of this still incurable neurodegenerative disorder.

Methods

Participants

The present study was based on data from the INSIGHT-preAD cohort¹⁵, a longitudinal study with a five-year follow-up of 318 older adults aged 70–85 years, presenting subjective memory complaints, but normal cognition, as defined by a Mini-Mental State Examination (MMSE) score ≥27²¹, a Clinical Dementia Rating (CDR) of 0⁵², and a Free and Cued Selective Reminding Test (FCSRT) total recall score ≥41²⁰. Participants were excluded if they had objective cognitive impairment or any neurological or psychiatric disorder that could affect cognition. Additional exclusion criteria included major medical conditions or sensory impairments that could interfere with participation or cognitive assessment, illiteracy, institutionalisation or legal guardianship, and contraindications to neuroimaging procedures such as MRI or PET scanning.

The study protocol was approved by the local ethics committee of Pitié-Salpêtrière University Hospital and was conducted in accordance with the principles of the Declaration of Helsinki. All participants provided written informed consent after receiving an explanation of the study two weeks prior to enrolment. All ethical regulations relevant to human research participants were followed.

At baseline, participants were stratified by amyloid status on 18F-florbetapir PET (A+, amyloid-positive; or A−, amyloid-negative) and underwent assessments of demographics, cognition and psychobehaviour characteristics, APOE ε4 status, brain structure and function (MRI), brain glucose-metabolism on 18F-fluorodeoxyglucose (18F-FDG) PET, and EEG. Both participants and researchers were blind to participants’ amyloid status. Follow-up included clinical and cognitive evaluations every six months, neuropsychological testing and EEG every year, and MRI and PET every two years.

For amyloid PET images, the threshold for normal (amyloid-negative) vs. abnormal (amyloid-positive) uptake was set to 0.7918⁵³. Of the 318 participants, 88 individuals were A+ at baseline. Among these, 15 developed clinical signs of AD during follow-up, showing cognitive decline with consistent episodic memory impairment observed at two consecutive visits, fulfilling the diagnostic criteria proposed by the International Working Group⁵⁴. At this time point (Mdiag), they were all at a prodromal stage. The diagnosis of AD was determined by an independent committee of two neurologists, a neuropsychologist, and a neuroimaging expert.

In addition to these 15 A+ progressors, who constitute the primary group of interest of the current study (PROG group), we randomly selected two control groups of 15 individuals each who remained cognitively stable throughout the five-year study period: amyloid-positive (CTRL A+ group) or amyloid-negative (CTRL A– group). The three groups were matched in terms of age, sex, and education level. Table 1 shows the demographic and clinical characteristics of the three groups at baseline (month 0, M0) and at the last visit. For the control groups, the last visit corresponds to month 60 (i.e. Mlast), whereas for the PROG group it corresponds to the visit at which the diagnosis was established (i.e. Mdiag). Current analyses do not include neuroimaging or electrophysiological data. Full INSIGHT-preAD cohort characteristics and methodological details are reported in ref.¹⁵.

Word recognition memory task

Every year, participants completed a word recognition memory task, as part of a larger protocol fully described in refs. ^10,15. The stimuli consisted of 16 target words sampled from the FCSRT²⁰ and 64 distractor words that matched the target words in frequency of occurrence and length, presented in four successive blocks of 52 trials each (208 trials in total). Lists A and B of the FCSRT were alternated each year. Figure 3 illustrates the experimental procedure and the resulting response categories. Specifically, the word memory recognition task consisted of two questions. The first question (Q1) assessed individuals’ ability to recognize whether a word presented on a computer screen had been previously seen, following administration of the FCSRT by a neuropsychologist one to three hours earlier. The second question (Q2) corresponded to a second-order judgment assessing online performance monitoring via individuals’ confidence in their Q1 performance by asking: “Are you sure of your response?”

Fig. 3. Experimental procedure for the word recognition memory task.

Participants first answered Q1, indicating whether they had previously seen the word (target/old item or distractor/new item, R1), and then answered Q2, indicating whether they were sure of their response (R2). These second-order judgments were used to assess online performance monitoring, as a proxy for subjective awareness; however, they do not provide a specific measure of awareness per se and may also reflect underlying memory strength. R1 and R2 together defined the self-awareness-related response categories (Correct Sure, CS; Correct Not Sure, CNS; Error sure, ES; Error Not Sure, ENS). ES and CNS constitute the key factors of the self-monitoring inefficiency index (SMI).

Responses to Q2 were then categorized as “Sure” (high confidence, reflecting certainty in Q1) or “Not sure” (low confidence, reflecting uncertainty in Q1). Combining these confidence judgments with the accuracy in Q1 produced four possible response patterns: i) Correct Sure (CS): correct response with certainty; ii) Correct Not Sure (CNS): correct response with uncertainty; iii) Error Sure (ES): erroneous response with certainty (indicative of error unawareness); iv) Error Not Sure (ENS): erroneous response with uncertainty.

Analytical approach

To evaluate five-year trajectories of response awareness and error-monitoring, we derived two specific indices from participants’ responses across annual visits: the Task Performance Index (TPI) and the Self-Monitoring inefficiency Index (SMI).

Specifically, to assess participants’ performance abilities in the word recognition memory task, we calculated the TPI, defined as the difference in accuracy between the last and first visits (accuracy at Mlast/Mdiag minus accuracy at M0) for each individual. Positive TPI values indicate performance improvement, whereas negative values indicate decline. TPI scores were then compared across groups to characterize differences in performance trajectories. Finally, associations between TPI and changes in SMI from the first to the last visit were examined.

Additionally, to study the impact of self-monitoring efficiency on task performance over time, we computed the SMI, derived from two key measures of impaired metacognitive monitoring — ES and CNS —, defined as: SMI = (ES + CNS) / total number of trials. Both ES and CNS reflect mismatches between objective task performance and subjective confidence, which constitutes the core components of error-monitoring, with confidence judgments and error monitoring likely relying on overlapping mechanisms. Expressed as a proportion of total trials, the SMI captures an individual’s overall load of poor confidence judgments independently of performance level at each visit. Correlations between the SMI and task accuracy over time were conducted in each group to assess how error-monitoring inefficiency may shape cognitive trajectories over time. Last, we investigated which specific error-monitoring factor (ES or CNS) primarily accounted for the group differences in performance changes.

Statistics and reproducibility

All analyses were performed in R (version 4.3.2; R Core Team, 2023). Group comparisons for demographic, clinical, and behavioral measures used non-parametric tests: Fisher’s exact test for categorical variables, and the Kruskal–Wallis test with Dunn’s post hoc tests (Holm correction) for continuous variables. Within-group comparisons between M0 and Mlast/Mdiag were performed using the Wilcoxon signed-rank test.

Trajectories of erroneous and correct responses, distinguishing “Sure” and “Not sure” judgments, were modeled using two separate Poisson generalized linear mixed models (GLMMs), one for erroneous and one for correct responses, with each model fitted using lme4 (v1.1-35.1). Group, Time (i.e., visits M0, M12, M36, M48, M60, treated as a continuous variable), Certainty Pattern (Sure vs. Not sure), and all interaction terms were specified as fixed effects, and participant ID as a random intercept. The significance of main effects and interactions, including the three-way interaction (Certainty × Time × Group), was assessed using Type II Wald chi-square (χ²) tests from the car package (v3.1-2). Annual rates of change were estimated using simple slopes and 95% confidence intervals derived with emtrends() from emmeans (v1.8.9), providing group- and certainty-specific estimates of temporal changes. Differences between Sure and Not sure patterns were additionally compared across groups at M0 and Mlast/Mdiag, with estimated marginal means (emmeans) differences ± standard errors, and diff-in-diff comparisons depicting group-specific temporal dynamics. All emmeans pairwise comparisons used Tukey adjustment for multiple comparisons.

To examine how task performance accuracy relates to the Self-Monitoring inefficiency Index (SMI) across groups and visits, we fitted a linear mixed-effects model including SMI, Group, Visit (categorical, M0 to M60), and all interactions, with participant ID as a random intercept. Type II Wald χ² tests assessed fixed effects, and post hoc trend analyses were performed using emmeans with Tukey adjustment for pairwise comparisons. To assess the distinct contributions of the two SMI components, ES and CNS, a set of linear mixed-effects models was fitted within the PROG group. All models included Visit as a fixed effect and participant ID as a random intercept. We compared a baseline Visit-only model with models including ES, CNS, or both predictors, and additionally tested ES × Visit and CNS × Visit interactions. Nested models were evaluated using likelihood-ratio tests, and the relative contributions of ES and CNS were quantified using log-likelihood differences and unique marginal R² values⁵⁵, as implemented in the MuMIn package (v1.46.0). Finally, associations between TPI and changes in SMI between first and last visits were examined using Spearman correlations. All tests were two-sided, with statistical significance defined as p < 0.05 after adjustment when applicable.

All analyses were conducted on predefined cohorts (n = 15 per group) and were fully reproducible using the reported R packages and versions, with repeated measures defined at the participant level.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

K.A. conceptualized and conducted the study, participated in the acquisition of clinical data as a medical investigator of the INSIGHT-preAD cohort, interpreted results, and wrote the first draft and final manuscript. F.X.L. performed statistical analyses, interpreted results, prepared figures and tables, and wrote the first draft and final manuscript. N.H. contributed to methods, data interpretation, and wrote the final manuscript. R.Y. contributed to methods and participated in editing the final version of the manuscript. M.T.S. contributed to study conceptualization and discussion of results. V.P. contributed to study conceptualization, data interpretation, and wrote the first draft and final manuscript.

Peer review

Peer review information

Communications Biology thanks George Prigatano, Thomas Benke and Jose Manuel Valera Bermejo for their contribution to the peer review of this work. Primary Handling Editors: Jessica Peter and Jasmine Pan. A peer review file is available.

Data availability

The current study was based on data from the INSIGHT-preAD cohort, whose principal investigator is Prof. Bruno Dubois. Study data are available at https://www.gaaindata.org/partner/INSIGHT-preAD. The data supporting both Figs. 1 and 2 are provided in a Supplementary Excel file.

Code availability

The complete datasets and analysis code that support the findings of this study can be shared upon reasonable request from the corresponding author [K.A.].

Competing interests

M.T.d.S. is an Editorial Board Member for Communications Biology, but was not involved in the editorial review of, nor the decision to publish this article. The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s42003-026-10461-z.