Personalized targeted memory reactivation enhances consolidation of challenging memories via slow wave and spindle dynamics

AI summary

Overview

Research area: Sleep-dependent memory consolidation, targeted memory reactivation (TMR), and EEG-based neurotechnology (submitted under Human-Computer Interaction, cs.HC).

Technical level: Advanced — the paper combines a controlled sleep electrophysiology experiment with event-related potential (ERP), time-frequency, phase-amplitude coupling, and multivariate SVM classification analyses.

Scope: The paper tests whether a TMR protocol that adjusts the number of auditory cue repetitions according to each person's own recall performance and self-rated task difficulty produces better memory consolidation than a uniform TMR protocol or no stimulation, and links any benefits to slow wave and sleep spindle dynamics during non-REM sleep.

What This Paper Is About

Targeted memory reactivation delivers learning-related cues during sleep to reactivate and strengthen memories, but standard protocols present the same cues to everyone in the same way. This ignores individual differences in learning ability and memory trace strength, which is thought to limit TMR's usefulness for memories that are hard to recall.

The authors built a personalized TMR protocol that varies how many times a cue is repeated based on whether the participant had recalled that item correctly and on how difficult they expected that item to be. They then compared this against a uniform TMR group and a no-stimulation control group in a word-pair memory task, measuring both memory performance and sleep EEG.

Key Contributions

1. A personalized cueing protocol. During sleep, the personalized TMR group received 1 presentation (PRES) for items rated L2 that were answered correctly, 2 PRES for L2 items answered incorrectly, and 4 PRES for L3 items regardless of correctness. The uniform TMR group received the same stimuli for all word pairs, and the control (CNT) group received no auditory stimulation.

2. A three-group behavioral comparison across difficulty levels. Thirty-six participants (12 per group; 13 females; mean age 26.57 ± 3.13 years) learned 104 semantically related word pairs across a pre-sleep and post-sleep session, with stimulation (or no stimulation) during an overnight sleep session from 10 p.m. to 6 a.m.

3. Neural evidence of enhanced oscillatory coordination. EEG analyses of ERP, time-frequency representations, and event-related phase-amplitude coupling (ERPAC) showed that stimulation increased slow wave (SW) and spindle activity and SW–spindle coupling relative to no stimulation, with distinct patterns between the personalized and uniform protocols.

4. Identification of group-specific neural signatures. An SVM with an RBF kernel classified EEG-derived features (SW power, spindle power, SW–spindle coupling) by group at above-chance accuracy, and a positive correlation between behavioral improvement and SW–spindle coupling appeared only in the personalized TMR group under the hardest condition.

Main Findings

• Group matching. No significant group differences were found in demographics, questionnaire scores, or sleep architecture (Supplementary Table 1), so performance changes were attributed to the intervention rather than participant characteristics or sleep parameters.

• Difficulty ratings were valid. Prospective self-rated difficulty correlated strongly negatively with retrieval accuracy across all groups and sessions (rho = –0.878 to –0.948; all p < 0.001).

• All items combined. Significant main effects of time (F1,66 = 42.42, p < 0.001) and group (F2,66 = 5.10, p = 0.009) appeared, with no interaction (F2,66 = 0.74, p = 0.479). All three groups improved from pre- to post-sleep: personalized TMR (t11 = -10.22), TMR (t11 = -8.37), CNT (t11 = -6.93), all p < 0.001. Post-hoc comparisons showed no significant group differences within either session.

• L1 (easiest) items showed no consolidation benefit. A significant group effect appeared (F2,66 = 3.61, p = 0.033), but time (F1,66 = 1.85, p = 0.179) and interaction (F2,66 = 0.63, p = 0.538) did not, and no group changed significantly from pre- to post-sleep.

• L2 (moderate) items improved with stimulation. Group (F2,66 = 9.47) and time (F1,66 = 12.55) effects were significant (both p < 0.001), interaction was not (F2,66 = 1.32, p = 0.274). Personalized TMR (t11 = -3.73, p = 0.003) and TMR (t11 = -5.65, p = 0.001) improved; CNT did not (t11 = -1.41, p = 0.187). Personalized TMR outperformed CNT in the post-sleep session (t22 = 4.33, p < 0.001).

• L3 (hardest) items separated the protocols. Group (F2,66 = 16.54), time (F1,66 = 20.77), and interaction (F2,66 = 8.91) were all significant (all p < 0.001). Only personalized TMR improved (t11 = -5.68, p < 0.001); TMR (t11 = -1.50, p = 0.161) and CNT (t11 = -1.56, p = 0.148) did not. Personalized TMR outperformed both TMR (t22 = 3.96, p < 0.001) and CNT (t22 = 4.49, p < 0.001) post-sleep.

• Pre- to post-sleep change scores. One-way ANOVA showed significant differences in All (F2,33 = 3.83, p = 0.032), L2 (F2,33 = 4.00, p = 0.028), and L3 (F2,33 = 14.18, p < 0.001). Personalized TMR outperformed CNT in All (t22 = 2.69, p = 0.013); TMR outperformed CNT in L2 (t22 = 2.94, p = 0.008); personalized TMR outperformed both TMR (t22 = 3.89, p = 0.008) and CNT (t22 = 4.54, p = 0.002) in L3.

• Error correction improved. Across memory transitions (correct-correct, correct-incorrect, incorrect-correct, incorrect-incorrect), no group differences appeared for CC or CI. For incorrect-correct transitions, differences were significant in All (F2,33 = 4.13, p = 0.025) and L3 (F2,33 = 3.43, p = 0.044), with personalized TMR higher than CNT in All (t22 = 2.84, p = 0.010) and L3 (t22 = 2.63, p = 0.015). For incorrect-incorrect transitions, a difference appeared in All (F2,33 = 3.45, p = 0.044), with CNT higher than personalized TMR (t22 = -2.70, p = 0.013).

• ERP responses differed by group and region. Personalized TMR showed significantly negative amplitudes immediately after the second stimulus followed by positive amplitudes during the 3–4 s interval, while TMR showed negative amplitudes in the same interval. In L3, personalized TMR showed significantly greater neural responsiveness than both TMR and CNT. Within 4-PRES sequences in personalized TMR, the first and last cues differed significantly, suggesting neural adaptation across repetitions (Supplementary Fig. 1). Differences were strongest in frontal (F3, F4) and central (C3, C4) channels and minimal in occipital channels (O1, O2).

• Slow wave and spindle power increased with stimulation. TFR showed enhanced SW and spindle band activation in personalized TMR and TMR relative to CNT. In the 0–0.5 s post-stimulus interval, TMR had greater SW activation than personalized TMR; in the 2–4 s interval, personalized TMR showed pronounced increases in both SW and spindle power relative to CNT. Effects were strongest over frontal and central regions.

• SW–spindle coupling (ERPAC, 4–20 Hz) increased with stimulation. In All, personalized TMR and TMR showed higher ERPAC than CNT. In L3, ERPAC values were highest in TMR, with personalized TMR slightly lower, which the authors suggest may reflect auditory habituation from repeated stimulation.

• Key EEG feature comparisons. SW power differed by group in All (F2,33 = 6.52, p = 0.004) and L3 (F2,33 = 6.01, p = 0.006): personalized TMR > CNT (t22 = 2.72, p = 0.013 in All; t22 = 2.69, p = 0.013 in L3) and TMR > CNT (t22 = 3.79, p = 0.001 in All; t22 = 3.53, p = 0.002 in L3). Spindle power differed in All (F2,33 = 3.31, p = 0.049) and L3 (F2,33 = 3.34, p = 0.048), with personalized TMR > CNT in All (t22 = 2.77, p = 0.011) and L3 (t22 = 2.63, p = 0.015). SW–spindle coupling differed only in L3 (F2,33 = 5.02, p = 0.013), with both personalized TMR (t22 = 2.62, p = 0.016) and TMR (t22 = 2.86, p = 0.009) stronger than CNT.

• Behavior–EEG correlation was specific and narrow. In L3, only the personalized TMR group showed a significant correlation: accuracy difference was positively associated with SW–spindle coupling (r = 0.70, p = 0.011). Correlations with SW power and spindle power were not statistically significant, and no significant correlations appeared in the All condition in any group.

• Classification results. Each EEG feature was independently classified with an SVM using an RBF kernel, exceeding chance level. Accuracy was higher after the second stimulus (2 s) than the first (0 s) and peaked around 3 s. Surrogate decoding used 250 label shuffles, and statistical significance was assessed with a two-sided cluster-based permutation test using 1,000 randomizations.

Methodology in Plain English

The study ran an overnight, three-group experiment with three sessions: pre-sleep, sleep, and post-sleep.

Participants. Forty-two healthy adults with normal hearing and vision were recruited and told to avoid caffeine, alcohol, and central nervous system-active drugs for 24 hours beforehand. They were randomly assigned to personalized TMR, TMR, or CNT with 14 per group. Six were excluded: two each from personalized TMR and CNT due to device malfunctions, and two from TMR as outliers (sleep efficiency below 70% and excessive wake after sleep onset). The final sample was 36 (12 per group). The study was approved by the Korea University Institutional Review Board (KUIRB-2022-0222-04), and participants gave written informed consent per the Declaration of Helsinki.

Preparation. One week before the lab visit, participants completed the Pittsburgh Sleep Quality Index, Insomnia Severity Index, and Self-Rating Depression Scale. On arrival at 6 p.m., they completed the Stanford Sleepiness Scale and State-Trait Anxiety Inventory, then did resting-state recordings and the memory task.

Memory task. Participants memorized 104 semantically related word pairs (for example, "event–festival") in randomized order. Each encoding trial began with a 1-s fixation cross, then a 4-s audiovisual word pair: the pair was shown on screen for 4 s while the upper word was played at stimulus onset (0 s) and the lower word at 2 s, each auditory word lasting roughly 300–900 ms depending on phonological length. In retrieval, the upper word was presented visually and auditorily as a cue, and participants had 10 s to type the lower word. After each retrieval trial they rated expected recall difficulty on a three-level scale: L1 (little difficulty immediately and after 12 h), L2 (easy immediately but challenging after 12 h), and L3 (significant difficulty both immediately and after 12 h). These ratings were prospective, based on self-evaluation rather than accuracy, and were not updated or reclassified later. In the pre-sleep retrieval phase only, each trial was followed by a 1-s break and a 4-s re-encoding period regardless of accuracy. Tasks ran in Psychtoolbox 3.

Sleep and stimulation. Participants slept from 10 p.m. to 6 a.m. with polysomnography. A sleep expert monitored EEG in real time and delivered auditory cues when participants had been stable in NREM 2 or NREM 3 for at least 10 consecutive 30-s epochs; stimulation paused if wake, REM, or NREM 1 was detected and resumed on return to stable NREM 2 or NREM 3. Each cue was a complete word pair presented sequentially within a 4-s window at approximately 45 dB, with a 4-s inter-stimulus interval; because the second word typically ended around 2.3–2.9

Authors’ abstract

Sleep is crucial for memory consolidation, underpinning effective learning. Targeted memory reactivation (TMR) can strengthen neural representations by re-engaging learning circuits during sleep. However, TMR protocols overlook individual differences in learning capacity and memory trace strength, limiting efficacy for difficult-to-recall memories. Here, we present a personalized TMR protocol that adjusts stimulation frequency based on individual retrieval performance and task difficulty during a word-pair memory task. In an experiment comparing personalized TMR, TMR, and control groups, the personalized protocol significantly reduced memory decay and improved error correction under challenging recall. Electroencephalogram (EEG) analyses revealed enhanced synchronization of slow waves and spindles, with a significant positive correlation between behavioral and EEG features for challenging memories. Multivariate classification identified distinct neural signatures linked to the personalized approach, highlighting its ability to target memory-specific circuits. These findings provide novel insights into sleep-dependent memory consolidation and support personalized TMR interventions to optimize learning outcomes.