Mnemonic-trained brain tuning to a regular odd-even pattern subserves digit memory in children

The study investigates how mnemonic training enhances memory in children and what neural mechanisms underlie such benefits. Two competing hypotheses are tested: (1) the processing-efficiency hypothesis, predicting reduced neural activity after training due to more efficient encoding; and (2) the encoding-pattern hypothesis, predicting a change in how items are encoded, specifically a differential response to odd- versus even-position digits in sequences when using a digit-image mnemonic. Prior work with superior mnemonists suggests that digit pairs (formed at even positions) are mapped to images, leading to longer processing at even positions and potentially distinct neural signatures (e.g., P200 differences). The authors formalize four models: Model A (efficiency: overall decreased neural activity post-training), Model B (pattern: emerging odd-even differences post-training), Model C (both efficiency and pattern changes), and Model D (no changes). The study aims to determine which model best explains training-induced changes in neural dynamics during encoding and whether such changes relate to memory improvement.

The paper reviews evidence that mnemonic strategies (e.g., Method of Loci, digit-image) can substantially improve memory across ages and are associated with changes in functional brain organization. The processing-efficiency account is grounded in expertise research showing reduced neural activity with improved performance in domain-specific tasks. Conversely, studies of superior mnemonists using digit-image methods indicate segmentation of digit sequences into pairs mapped to images, with longer encoding times at even positions, suggesting an encoding-pattern bias. Preliminary EEG work showed higher P200 for even-position digits. P200 has been linked to top-down attention during mental imagery, and theta oscillations are robustly implicated in memory encoding and mental imagery, predicting subsequent memory and reflecting cognitive effort. The authors also note prior training studies in children reporting neural changes but lacking clarity on encoding-stage dynamics during mnemonic use.

Design: Longitudinal study with four sessions: pre-training (EEG task and cognitive battery), 22-day training period, post-training (EEG task and cognitive battery), and a 4-month follow-up (subset of cognitive tests). Participants: 41 children (mean age ~13 years; 22 boys) recruited from schools in Hainan, China; no prior memory training. Assignment to mnemonic training (MT; n=20) or no-contact control (NC; n=21) based on interest. Exclusions due to EEG artifacts, low task performance, or apparatus issues resulted in 33 participants for analyses (MT n=16; NC n=17). Ethical approval obtained; consent provided. Mnemonic training: MT group attended a World Memory Championships Training Camp (22 days, ≥6 h/day). They learned a 00–99 digit-image conversion table (e.g., 72→penguin via phonological mapping in Chinese; 51→worker via semantic association) and practiced multiple competition disciplines (Spoken Numbers, Binary Digits, Hour Numbers, Random Numbers) and everyday digit memorization. EEG memory task: 20 blocks. Encoding: 12 digits presented sequentially; each trial started with fixation (0.5 s), digit display (2 s), ISI 0–0.5 s. Retention: 6 s blank. Retrieval: fixation (0.5 s), 2-digit probe (1 s), response screen (Y/N) up to 3 s. Probes were 2-digit combinations from odd–even serial positions within the studied sequence (6 probes per block; half targets). Performance measured by d' (z(hit)−z(false alarm)). Cognitive battery: Perceptual speed (two-choice reaction: digit, figure), inhibitory control (Stroop color), working memory (2-back digital, 2-back spatial), short-term memory (digit matrix), episodic memory (word lists; number–noun pairs), reasoning (grammatical reasoning), spatial imagination (Purdue rotations), divided attention (multi-object tracking). Follow-up retested digit matrix and number–noun pairs. EEG acquisition: 64-channel cap (10/10 system), Neuroscan Synamps2; V/H EOG recorded; impedance <10 kΩ; sampled at 500 Hz; online 0.05–100 Hz bandpass; average mastoid reference. Preprocessing: EEGLAB and MATLAB. Downsampled to 250 Hz; epochs −500 to 2000 ms from digit onset; ICA to remove eye artifacts; re-referenced to common average; baseline −500–0 ms; artifact rejection ±100 µV; ERP low-pass 40 Hz. Trials sorted by Position (even vs odd). On average per condition: ERP ~90.4 trials (range 71–118); oscillatory ~84.2 trials (range 62–108). ERP analysis: Focus on P200 (170–220 ms). Electrodes grouped into nine Areas (left/medial/right anterior; left/medial/right central; left/medial/right posterior). Planned repeated-measures ANOVAs run separately for MT and NC with within-subject factors Position (even vs odd), Session (pre vs post), Area (9 levels). Bonferroni correction for post-hoc comparisons. Oscillatory power: ERSP with 5-cycle Morlet wavelets, 4–30 Hz (27 linear steps). Power expressed as percent change from −500–0 ms baseline. Focus on frontal theta (4–8 Hz) averaged over 200–1000 ms. Frontal Areas: left anterior, medial anterior, right anterior (specified electrode sets). Planned three-way ANOVAs (Position × Session × Area) per group with Bonferroni-corrected post-hocs. Validation analyses for processing-efficiency: Cluster-based ERP comparison (0–2000 ms) between pre and post within MT (FDR-corrected, ≥10 consecutive time points and ≥4 adjacent electrodes). Cluster-based ERSP comparison across theta (4–8 Hz), alpha (9–12 Hz), beta (13–30 Hz) bands to detect general pre–post differences. Statistics: Behavioral EEG task d' analyzed via two-way mixed ANOVA (Group × Session). Cognitive battery measures analyzed via mixed ANOVAs with FDR correction across tests. Neural–behavioral correlations: Pearson correlations between memory increment (d'post−d'pre) and post-training even−odd P200 effect (averaged significant electrodes) and even−odd theta effect (e.g., F3), with Fisher's z to compare groups.

Behavioral (EEG task): Significant Group × Session interaction, F(1,31)=15.46, p<0.001, η²partial=0.30. MT improved from pre (d'=1.37±0.12) to post (2.10±0.19), p<0.001; NC showed no change (pre 1.12±0.15; post 1.16±0.10), p>0.05. Cognitive transfer: After FDR correction, only number–noun pairs showed a significant interaction indicating superior improvement in MT vs NC that persisted at follow-up. No other measures showed significant interactions, indicating limited generalization. ERP (P200, 170–220 ms): In MT, main Position effect F(1,15)=8.74, p<0.01, η²partial=0.33; Position × Area F(8,120)=3.79, p<0.05; critical Position × Session interaction F(1,15)=3.79, p<0.05. Post-training, P200 amplitude was greater for even vs odd positions (even 3.08±0.96 µV vs odd 1.34±0.80 µV, p<0.05). No significant Session main effect. In NC, no significant effects involving Session or Position (all ps>0.41). Oscillatory power (theta 4–8 Hz, 200–1000 ms): In MT, Position main effect F(1,15)=5.40, p<0.05, η²partial=0.31; Area × Position F(2,30)=3.82, p<0.05; three-way interaction F(2,30)=6.92, p<0.05, η²partial=0.37. Post-training showed Position × Area interaction F(2,30)=14.43, p<0.01; left anterior theta increased from odd (10±5%) to even (24±6%), p<0.05. Pre-training showed no such effect. NC showed no significant results (all ps>0.09). Neural–behavioral correlations: In MT, post-training even−odd P200 effect correlated with d' improvement (r=0.66, p=0.009); correlation larger than NC (r=0.13, p=0.63), Fisher’s z=1.72, p=0.04. Left anterior frontal theta effect at F3 correlated with d' improvement in MT (r=0.52, p=0.04) but not NC (r=0.06, p=0.82); between-group difference z=1.34, p=0.09. Validation analyses: No significant global ERP pre–post clusters. Alpha band (9–12 Hz) power increased post vs pre in 1000–1400 ms window (post 18±2% vs pre 6±3%, Pcorr<0.05) but did not correlate with memory increment (r=0.23, p=0.57). Overall, findings align with Model B (encoding-pattern change) rather than a general processing-efficiency decrease.

The results support the encoding-pattern hypothesis: after digit-image mnemonic training, children exhibited a regular odd–even neural pattern during encoding—enhanced P200 and increased frontal theta for even-position digits relative to odd positions—corresponding to the strategy of pairing digits into images at even positions. These neural changes emerged specifically post-training and predicted individual gains in memory performance, indicating functional relevance. The timing suggests both early attentional allocation (P200 around 170–220 ms) and sustained processing (theta 200–1000 ms) contribute to reorganized encoding. The pattern is consistent with organizational and chunking mechanisms observed in expert memorizers and with frontal involvement in imagery and working memory maintenance. Evidence for general processing-efficiency (global neural decreases) was limited; although alpha power increased post-training, it did not relate to performance gains, and no broad ERP reductions were detected. Thus, mnemonic training primarily reshaped the encoding pattern rather than broadly reducing neural effort. The limited transfer to untrained cognitive domains aligns with domain-specificity of mnemonic expertise and suggests trainees may not spontaneously apply strategies beyond trained tasks.

Mnemonic training using a digit-image method enhanced short-term digit memory in children and induced a distinctive odd–even encoding pattern in the brain, reflected by increased P200 amplitude and frontal theta power for even-position digits. These neural signatures predicted individual improvements in memory, providing electrophysiological evidence that mnemonics support memory by altering encoding patterns (consistent with Model B). General processing-efficiency reductions were not clearly supported. Future research should: (1) increase sample sizes and statistical power; (2) examine network-level efficiency (e.g., phase synchrony, small-world properties) to test processing-efficiency mechanisms; (3) compare different mnemonic strategies and their neural signatures; and (4) investigate methods to promote transfer of mnemonic skills to broader cognitive domains.

The study had a modest sample size without an a priori power analysis, potentially limiting statistical power. Group assignment was based on participant interest, introducing potential selection biases and background differences (e.g., motivation), which may limit generalizability despite the within-subject pre–post design. Planned-contrast EEG analyses focused on within-group effects, and the absence of robust Group interactions reduces conclusiveness regarding between-group differences. Evidence for processing-efficiency might exist at the network level or below detection thresholds of the current ERP/ERSP analyses. Transfer effects to untrained domains were limited, possibly because trainees did not spontaneously apply the mnemonic outside the trained context.