Human neural dynamics of real-world and imagined navigation

The study asks whether human medial temporal lobe (MTL), particularly hippocampal, dynamics that support spatial navigation also organize internally generated sequences during imagination. Prior rodent work shows hippocampal place cells and theta-coordinated sequences persist during immobility and support episodic memory and planning. In humans, theta oscillations occur in brief bouts rather than continuously, raising uncertainty about their capacity to support sequences and event segmentation in real-world navigation or imagination absent external cues. The authors aim to compare MTL theta dynamics during real-world ambulatory navigation and explicitly instructed imagined navigation of the same routes, testing whether intermittent human theta bouts encode route structure and relative position, and whether such dynamics generalize to imagination.

Prior research indicates the MTL and hippocampus are central to episodic memory and navigation. In rodents, place cells form sequences that span movement trajectories and persist during immobility, supporting memory organization and planning. Theta oscillations (~4–12 Hz) coordinate these sequences; their disruption impairs spatial memory. Human hippocampal theta appears in intermittent bouts during virtual and real-world navigation, unlike the continuous theta in rodents, leaving open whether human theta bouts can support segmentation and internally generated sequences. Related work shows hippocampal dynamics encode positions, times, and sequences and may mark event boundaries, suggesting a shared organizational framework across navigation and memory. The study builds on these findings to test for shared neural structure between real and imagined navigation in humans.

Participants: Five adults (24–40 years; 3 males, 2 females) with pharmaco-resistant focal epilepsy chronically implanted with the NeuroPace RNS System volunteered. Electrode placement was determined clinically; recordings were selected from medial temporal lobe (hippocampus, parahippocampal, perirhinal, subiculum). Across participants, 18 bipolar MTL channels were analyzed. Informed consent and UCLA IRB approval were obtained.

Tasks and design: Participants learned two indoor walking routes (room 14.6 × 13.5 m) with five linear segments and four turns (left and right variants). Learning used tablet-displayed ideal route shapes and motion-capture feedback. Six life-size object cutouts (three per route) were present during learning; during the third segment, an auditory cue prompted a self-location judgment (closest object) via handheld mouse. After learning, objects were removed. Real-world walking trials were interleaved with treadmill walking in three blocks: Block 1 control treadmill walking (no imagination); Blocks 2 and 3 imagination treadmill walking (participants mentally navigated their previous or upcoming route while walking at steady speed). Across the experiment: 144 walks (36 left, 36 right real-world; 24 control treadmill; 24 imagined left; 24 imagined right; imaginations split into previous and upcoming). Gaze was instructed to remain forward. Participants indicated imagination completion with a button press.

Behavioral acquisition: Motion capture (OptiTrack; 30 IR cameras; 120 Hz) tracked lower-body skeleton (Helen Heyes markers), head rigid-body pose; signals smoothed with 0.2 s Gaussian. Speed, head and hip yaw angular velocities derived. Eye tracking (Pupil Labs Invisible; 200 Hz eye streams; scene at 30 Hz) provided saccade detection (ClusterFix) and gaze density maps. A Unity 2021.2.19f1 application coordinated task control, synchronized motion capture, and logged behavioral responses.

iEEG acquisition: RNS-320 at 250 Hz, 1–90 Hz bandwidth, two depth leads per participant with four contacts (10 mm spacing); up to four bipolar channels per participant in MTL were selected. iEEG synchronized to motion and eye tracking via NTP markers. Interictal epileptiform discharges were detected and excluded using amplitude thresholds; 3.8 ± 3.0% of data flagged; trials with >50% IED coverage were removed.

Signal processing: Time–frequency amplitudes (1–90 Hz) via continuous wavelet transform (analytic Morse; symmetry 3; time–bandwidth 60; 10 voices/octave). Channels were z-scored to normalize amplitudes. Individual theta ranges were identified per subject by spectral peaks within 3–12 Hz bounded by neighboring minima. Additional delta (1–3 Hz) and beta (13–30 Hz) bands were examined based on spectral peaks.

Alignment: Real-walk trials were aligned using dynamic time warping to match turn timings; piecewise linear warping was applied to motion and time–frequency amplitudes. For imagination and control treadmill, trials were linearly warped to their condition-average duration; for imagination, durations were defined from treadmill onset to button press; later a model-based alignment to route structure with up to ±2 s shifts was learned using cross-validated dynamic time warping.

Theta bout detection: Transient oscillations were detected with eBOSC, requiring at least two cycles and power above the 95% CI at frequency-specific thresholds. Prevalence, duration, amplitude, and time-resolved bout rate (percent trials with a bout at each time) were computed.

Temporal consistency: Within each condition (real, imagined, control) and channel, trials were randomly split 1,000 times; the correlation between mean signals of halves quantified temporal consistency. Group comparisons used multilevel block permutation tests (participant as exchangeability block) with cluster bootstrap CIs; linear mixed-effects models assessed spatial correlations across channels between conditions.

Position encoding model and reconstruction: Route structure was abstracted as a sinusoid peaking prior to each turn; relative position within each segment was treated as a circular variable. Theta dynamics across all 18 MTL channels served as predictors in a linear regression to estimate cosine and negative-sine components corresponding to route phase. Tenfold cross-validation repeated 10 times prevented overfitting. Model trained on real-walk data reconstructed relative position within segments; performance assessed via normalized probability densities across actual vs estimated positions and circular errors (angle differences). For imagination, the real-trained model produced initial position estimates; dynamic time warping aligned these to the route structure with ±2 s maximum shift per participant; alignment learned in cross-validation and applied to held-out imagination trials. Identical procedures were applied to control treadmill trials.

Statistics: Nonparametric permutation and cluster-based permutation tests assessed significance of theta amplitudes aligned to turns and temporal consistency differences; linear mixed-effects modeled spatial correlations; regression analyses related theta to behavioral variables; cross-correlations quantified timing relationships; null distributions constructed via label permutation and temporal shifts for reconstruction errors and correlations to route structure.

• Participants learned and reliably executed two multi-turn routes; theta (3–12 Hz) amplitudes in MTL increased before turns, peaking significantly prior to physical turning (Δt ≈ −0.94 s; 95% CI −1.05 to −0.80; P<0.001; d=−1.09). Cluster-based permutation confirmed significant pre-turn theta increases (P<0.001), preceding speed deceleration and other behavioral changes.
• Theta bouts were intermittent (mean prevalence 21.2 ± 6.6%; mean duration 0.524 ± 0.077 s) and aligned to upcoming turns across trials. Trial-wise bout occurrence closely mirrored averaged theta amplitude dynamics.
• After learning, the prevalence of theta bouts preceding turns was higher than during learning (r_p=0.04; 95% CI 0.01–0.12; P=0.021; d=0.68).
• Theta dynamics depended on route segments and were consistent across all five participants and both routes; similar but distinct dynamics were observed in delta (1–3 Hz) and beta (13–30 Hz), with beta increases during the third segment coinciding with a self-location subtask.
• Temporal consistency of theta dynamics was highest during real-world navigation (r_a=0.26; 95% CI 0.22–0.36; P<0.001; d=3.06), elevated during imagined navigation versus control treadmill walking (r_a=0.15; 95% CI 0.09–0.26; P<0.001; d=1.83), and absent in control treadmill walking (P=0.964). A modest session effect suggested learning (r_a=0.18; 95% CI 0.12–0.24; P=0.023).
• Spatial congruence across MTL: channels with high temporal consistency during real-world navigation also showed high consistency during imagined navigation (linear mixed-effects, r=0.67; β=0.75; df=16; 95% CI 0.25–1.24; P=0.006), notably in the left anterior hippocampus.
• Delta-band temporal consistency increased for real and imagined navigation vs control (real r_a=0.29; 95% CI 0.20–0.38; P<0.001; imagined r_a=0.14; 95% CI 0.03–0.25; P=0.032), whereas beta showed increases during real but not imagined navigation.
• Position reconstruction from theta dynamics: The model reconstructed relative segment positions during real-world navigation above chance (δ=0.63; 95% CI 0.61–0.64; P<0.001), with estimates accurate across entire segments, especially near turns.
• Applying the real-trained model to imagination, after alignment, yielded significantly smaller reconstruction errors than control treadmill walking (Δδ = −0.43; 95% CI −0.46 to −0.40; P<0.001); control treadmill reconstructions failed.
• Theta dynamics correlated between real and imagined navigation (r=0.30; 95% CI 0.28–0.31; P=0.003) and with the abstracted route structure (real r=0.51; 95% CI 0.49–0.52; P<0.001; imagined r=0.29; 95% CI 0.28–0.31; P=0.010); no correlation in control treadmill (r=−0.01; P=0.528).
• Single-trial regressions during imagination showed theta related to estimated relative position (β_d_=0.02; 95% CI 0.01–0.04; P=0.006; d=1.93), with no significant effects of head rotation (P=0.051), hip rotation (P=0.982), eye movement (P=0.568), or speed (P=0.92).

Findings demonstrate that intermittent theta oscillations in the human MTL, particularly hippocampus, segment continuous ambulatory behavior into discrete route sections by aligning before turns, which act as task-relevant event boundaries. The replication of these structured theta patterns during imagined navigation—despite identical physical behavior to control treadmill walking and absence of external spatial cues—indicates that hippocampal networks can internally generate dynamics reflecting task structure and relative position. The shared spatial profile of temporal consistency across MTL sites during real and imagined navigation, strongest in the left anterior hippocampus, supports common anatomical substrates. Position reconstruction from theta dynamics generalized from real-world to imagined navigation after temporal alignment, showing that MTL theta encodes abstract route phase rather than solely physical kinematics. These results bridge research on navigation and episodic memory by showing that human hippocampal theta bouts can support internally generated sequences and event segmentation, consistent with rodent data on theta-coordinated sequences and event boundaries. The observed delta and beta activities suggest additional frequency-specific roles, with delta reflecting slower structured dynamics and beta potentially linked to self-location subtask or interactions between theta-driven inputs. Overall, the work supports a common neural organization for real-world and imagined navigation and suggests that hippocampal sequence organization extends to memory retrieval and future simulation.

Transient theta oscillations in the human MTL partition complex navigational routes into discrete segments and exhibit strikingly similar temporal structure during real and imagined navigation. A model trained on real-world theta dynamics accurately reconstructed relative position and generalized to imagination, indicating that hippocampal networks represent abstract route structure independent of sensory input. These findings support the idea that memory mechanisms leverage circuitry evolved for spatial navigation and provide a framework for studying internally generated sequences in episodic memory and prospective thought. Future work should combine single-neuron and local field potential recordings in freely moving humans to resolve neuronal generators of theta bouts, test generalization to non-spatial tasks with analogous temporal structure, clarify relationships between MTL theta and scalp EEG rhythms, and assess how novelty and goal states modulate theta segmentation during imagination.

• Small sample (n=5) of participants with epilepsy implanted with RNS devices; generalizability to broader populations requires confirmation.
• Electrode placements were clinically determined and varied across participants, limiting fine-grained subregional comparisons within the MTL beyond the observed left anterior hippocampal cluster.
• Imagined “future” routes were not novel but based on previously experienced trajectories; imagination and memory retrieval processes could not be fully disentangled.
• Imagination lacks direct measures of internal velocity and timing; alignment relied on model-based time warping, which assumes consistent structure across trials.
• Potential confounds from eye movements and movement kinematics were assessed and largely ruled out, but residual influences cannot be entirely excluded.
• Scalp EEG during movement had artifacts; primary conclusions are based on iEEG; relationships between MTL theta and scalp rhythms remain to be clarified.