Epilepsy is one of the most prevalent neurological disorders, affecting approximately 50 million people worldwide (Bell and Sander, 2001). While anti-seizure medications are the primary treatment, nearly one-third of patients develop drug-resistant epilepsy, creating a critical need for alternative therapeutic strategies (Kwan et al., 2011). For this population, deep brain stimulation (DBS) has emerged as an effective treatment, but its application is limited by the need for invasive neurosurgery, which carries inherent risks such as hemorrhage and infection (Lozano et al., 2019). Non-invasive brain stimulation, such as transcranial direct current stimulation (tDCS) and transcranial alternating current stimulation (tACS), has sought to provide safer alternatives, but these modalities have historically been limited by a trade-off between focality and depth, often resulting in diffuse stimulation of superficial cortical areas (Simula et al., 2022).

Temporal interference (TI) stimulation was introduced as a non-invasive alternative to overcome this limitation (Grossman et al., 2017). This novel technique leverages a physical principle that two high-frequency (>1 kHz) electric fields (called carrier frequencies) can generate low-frequency modulation transcranially. It has been hypothesized that the individual high-frequency fields oscillate too rapidly for neurons to follow, and instead pass through tissue with lower impedance than low frequency stimulation technology like tACS (<100 Hz). However, at their point of intersection, these high-frequency electrical fields interfere to create a low-frequency amplitude-modulated envelope that oscillates at the difference between the high frequencies. This low-frequency envelope can focally drive neural activity with high spatial precision, offering the hope of DBS-like targeting without invasive surgery (Grossman et al., 2017).

Since the original demonstration of temporal interference stimulation, multiple groups have expanded TI across theoretical, modeling, and in vivo systems. Subsequent studies have characterized the biophysical basis and focality of TI using finite-element modeling (Mirzakhalili et al., 2020), optimized electrode configurations and carrier frequencies (Cao and Grover, 2020; Stoupis and Samaras, 2022), and demonstrated modulation of hippocampal and cortical activity in rodent models (Acerbo et al., 2022; Missey et al., 2021). More recent work has extended TI to non-human primates and humans, highlighting both its translational potential and remaining safety considerations (Liu et al., 2024; Violante et al., 2023). Despite these advances, systematic evaluation of tissue-level safety markers under controlled TI conditions remains limited, motivating the present study.

However, safety concerns for TI stimulation stem not only from the high-frequency carriers but also from the low-frequency envelope they create. It is well-documented that low-frequency electrical currents can be harmful. Indeed, they are clinically used in radiofrequency thermocoagulation to intentionally create therapeutic lesions through resistive heating (Bourdillon et al., 2018). While the seminal work by Grossman et al. (2017) included an initial safety screen that found no evidence of apoptosis or DNA damage, it did not specifically investigate the thermal stress profile of the low-frequency envelope. This study directly addresses this critical knowledge gap. We hypothesized that while the low-frequency envelope generated by TI has the potential to cause thermal changes, its indirect generation via interfering high-frequency fields will produce a smaller thermal effect than the direct low-frequency stimulation.

This hypothesis is based on the distinction between a direct 5 Hz tACS current and the TI-generated 5 Hz envelope. tACS provides a true, low-frequency current that must overcome the high, frequency-dependent impedance of biological tissue (Z_HIGH), thereby generating significant heat P = I²Z (Gabriel et al., 1996). In this equation, I is the electrical current passing through the tissue, and P is the resulting electrical power released as heat (Joule heating). Because biological tissues act as capacitors, their capacitive impedance (Z) is inversely proportional to frequency (ω). Therefore, the low 5 Hz frequency of tACS encounters this Z_HIGH. The 5 Hz envelope generated by TI, in contrast, is not a current but rather a temporal pattern of amplitude modulation, which doesn’t follow the P = I²Z equation. Consequently, the only true currents propagating through the tissue are the high frequency carriers (>1 kHz). These carriers encounter a much lower tissue impedance (Z_LOW), and as a result, should produce minimal heat (Schwan, 1957). This physical distinction allows the 5 Hz neuromodulatory pattern to be effectively decoupled from the large thermal effects of a true 5 Hz current. To test this hypothesis, we performed a multi-modal safety assessment, including in vitro thermal-damage models, in vivo temperature recordings, and post-stimulation histological analysis.

In this multi-modal approach (Figure 1C), we first used an in vitro egg white model to conduct a proof-of-principle test, comparing the thermal coagulation effects of TI stimulation to conventional low-frequency tACS (Figure 1C top). We then moved to an in vivo mouse model to measure the intracranial temperature changes in the hippocampus during TI stimulation (Figure 1C top). Finally, a separate cohort of animals underwent the same stimulation paradigm and was sacrificed 24 h later for histological analysis to assess cellular markers for thermal stress (HSP70), inflammation (iNOS), and astrocyte activation (GFAP) (Figure 1C bottom). All stimulation waveforms were continuous and sinusoidal. TI was delivered using 1,000 Hz and 1,005 Hz carriers, selected to match kHz-range carrier frequencies validated in prior rodent TI studies and to remain within hardware limits. tACS was delivered through a single electrode pair because only one low-frequency source is required.

This study provides a multi-faceted safety evaluation of TI stimulation, specifically addressing the critical question of whether it induces thermal or cellular damage. Our convergent lines of evidence from in vitro, in vivo, and histological assessments consistently demonstrate that TI, under the parameters tested, exhibits a favorable short-term safety profile and that its neuromodulatory component is effectively decoupled from significant thermal injury.

The in vitro egg white experiment served as a proof-of-principle. While direct 5 Hz tACS caused visible protein coagulation, the 5 Hz TI envelope produced no such effect (Figure 2). This result provides initial support for our hypothesis that interfering high-frequency fields produces a smaller thermal effect than direct low-frequency stimulation.

It is important to explain the different current amplitudes used in our in vitro and in vivo models. The 10 mA current in the egg white model served as a highest amplitude to assess the physical limits of the TI principle, demonstrating that even at this high level, it did not cause thermal coagulation. This 10 mA level was not used in live mice because our preliminary testing confirmed it was biologically unsafe, inducing muscle twitching and the potential for off target effect. We therefore selected 2 mA as the highest amplitude that was well-tolerated by the animals for the in vivo safety study. This two-part approach allowed us to first confirm the physics of thermal decoupling in vitro (at 10 mA) and then assess the in vivo histological safety at a high but biologically tolerated amplitude (2 mA). Due to different impedance and size, transformation of the amplitude utilized on different model should be careful. Human TI studies typically apply 2–4 mA per electrode pair at kHz carrier frequencies and should be considered as a factor when interpreting the result (Missey et al., 2025; Violante et al., 2023).

This was subsequently supported in vivo, where we measured only a minor and stable temperature increase of 0.74°C ± 0.31°C (Figure 2D). This thermal load is well within the range of normal physiological fluctuations and is comparable to the<1°C increase reported for other non-invasive techniques like tDCS, which are widely considered safe under low amplitude for clinical use (Khadka et al., 2018). Although an earlier report shows no effect of hippocampal targeted TI stimulation on temperature in the cortex, our study suggests that TI induces minor temperature elevation, within safety ranges, in the target area (Grossman et al., 2017). However, since the animal underwent craniotomy in both control and experimental group, there can be modest difference in thermal and conductive properties that one should be aware.

The histological data further reinforce these findings. First, the complete absence of HSP70 expression in the hippocampus (Figure 4A) confirms the thermal increase was within the safety range, as HSP70 is a highly sensitive marker for temperature increases of 1°C–2°C (Yenari, 2013). Similarly, the lack of iNOS upregulation (Figure 4C) indicates that TI does not trigger a significant neuroinflammatory reaction. Second, our analysis showed no change in GFAP expression within the CA1 and DG subfields (Figures 3C, D). This finding is consistent with the safety screen by Grossman et al. (2017). However, A key contribution of our work is the identification of a novel, highly localized GFAP upregulation specifically within the SLM (Figure 3B), a subregion not a focus by Grossman et al. (2017) GFAP upregulation are often associated with astrogliosis, which is a canonical response to CNS injury (Eng et al., 2000). However, GFAP alone does not definitively establish reactive gliosis, which requires investigation among additional marker and functional assays.

Given the striking absence of corresponding heat shock or inflammatory markers, we argue against the isolated GFAP upregulation represents cellular damage. Instead, we hypothesize that this finding might reflect a circuit-specific physiological response, providing a possible marker for the preferential activation of the entorhinal-hippocampal perforant pathway by the TI stimulation. The SLM is a critical integration hub with extremely high density of synapses, and it is where the distal dendrites of CA1 neurons receive direct input from the entorhinal cortex via the temporoammonic pathway, a circuit fundamental to memory formation (Aksoy-Aksel and Manahan-Vaughan, 2013). This interpretation is strongly supported by recent human studies, which demonstrate that TI targeting the hippocampus also modulates entorhinal cortex activity and enhances memory, confirming functional engagement of this exact circuit (Violante et al., 2023). Therefore, the intense synaptic activity driven by TI in this input-heavy region could place a high metabolic demand on the local astrocytes, causing them to increase their activity. Thus, GFAP expression at the SLM acts more as a physiological, rather than pathological, response (Capogna, 2011; Dvorak-Carbone and Schuman, 1999). Lastly, an alternative, non-mutually exclusive hypothesis is that this finding might reflects functional astrocyte heterogeneity, suggesting SLM astrocytes are intrinsically more sensitive to electrical stimulation than those in other hippocampal layers (Khakh and Sofroniew, 2015). These interpretations remain speculative, and further experiments incorporating additional glial markers, functional assays, and larger cohorts will be required to establish mechanistic conclusions.

Interpretation of GFAP changes following electrical stimulation must also consider the heterogeneity of findings in the existing literature. While the original Grossman et al. (2017) study reported increased c-Fos and GFAP expression following TI, subsequent work has yielded mixed results. For example, Carmona-Barrón et al. (2023) reported inconsistent or reduced immediate-early gene expression following TI, and the preprint Peressotti et al. (2025) suggest that astrocytic responses may depend strongly on stimulation parameters, timing, and tissue context. Similar variability has been reported in studies of tACS and other electrical stimulation modalities (Carmona-Barrón et al., 2023; Monai et al., 2016; Peressotti et al., 2025). In addition, electrode implantation and craniotomy alone can induce local glial and stress responses, including GFAP and heat shock protein expression, which may confound interpretation of stimulation-related effects. While sham controls mitigate this concern, future non-invasive paradigms will be essential for translational relevance.

This study has several limitations. First, due to the small data size, the statistical power is limited, the findings should be interpreted as preliminary and will require confirmation in larger cohorts. Second, our findings are predicated on an acute, single-session stimulation paradigm. As chronic administration is required for most therapeutic applications, future preclinical studies must assess the cumulative effects of long-term TI. Third, our investigation utilized a single parameter set; a systematic exploration of different parameter space (e.g., amplitude, duration, envelope frequency) would fully delineate the therapeutic and safety window. Fourth, the observed increase in GFAP immunoreactivity within the SLM should be interpreted cautiously. GFAP alone does not definitively establish reactive astrogliosis, and one must be aware of potential GFAP and HSP70 changes due to surgical manipulation or electrode implantation. Finally, the isolated GFAP upregulation observed in the SLM, requires more mechanistic investigation. An expanded biomarker panel such as additional astrocytic markers (e.g., S100β), microglial markers (e.g., Iba1), cytokine profiling, and region-specific gene expression analysis could help confirmed the presence of potential localized astrogliosis and pathological stress.

Despite these limitations, our findings provide initial evidence that TI exhibit a favorable acute safety profile both in vitro and in vivo. TI produced only minimal temperature increases and no detectable induction of HSP70 or iNOS and the upregulation of GFAP was spatially restricted to the SLM.