Long-term optical monitoring of genetically encoded fluorescent indicators

Abstract

Over the past several decades, genetically encoded fluorescent indicators have revolutionized neuroscience by enabling cell-type-specific optical recording of neural activity. While most applications have focused on brain regions where stimulus-evoked activity correlates with behavior on the scale of seconds to minutes, many fundamental behavioral and physiological processes such as feeding, thermoregulation, and circadian timekeeping occur over hours to weeks. However, adapting optical recording techniques to these longer timescales presents unique challenges, particularly in accurately measuring and interpreting neural activity across extended recording durations. As a result, even studies using similar data have reached divergent conclusions, largely due to differences in data analysis and interpretation. This lack of standardization risks misinterpretation, miscommunication, and reduced reproducibility. In this article, we focus on in vivo fiber photometry calcium imaging in circadian neuroscience research as a case study. We review the current literature, outline theoretical, and practical challenges, and offer perspectives for optimizing experimental approaches and standardizing data interpretation. Importantly, the fundamental principles of long-term optical recording extend beyond circadian research and apply broadly to brain circuits that govern behavior and physiology over days to weeks.

Introduction

Circadian rhythms as a case study for long-term optical recording

Genetically encoded fluorescent indicators (GEFIs) have transformed systems neuroscience by enabling real-time monitoring of neural activity in genetically defined cell types. Recent advances now allow these signals to be recorded continuously for hours, days, or even weeks, offering new opportunities to study brain function across extended timescales (Table 1). As a result, researchers face new challenges and opportunities in designing, acquiring, and analyzing long-term recordings, particularly in systems where relevant neural signals evolve slowly over time.

Table 1.

Summary of published studies using long-term in vivo optical recording to monitor circadian neural and molecular rhythms.

Literature/year	Reporter	Brain area	Session length	Analysis method(s)	Isosbestic correction
S. Yamaguchi et al. 2001 (64 )	Per1-Luc	SCN	Continuous (binned to 5 min)	Mix (direct average)	N/A
Ono, Honma, and Honma 2015 (65)	Per1-Luc; Per2-Luc; Bmal1-Luc	SCN	Continuous (binned to 1 min)	Tonic Integration Peaks	N/A
Ono, Honma, and Honma 2015 (66)	Per2-Luc	SCN OB	Continuous (binned to 1 min)	Mix (direct average)	N/A
Y. Yamaguchi et al. 2016 (67)	Per1-Luc; Per2-Luc	SCN (rat)	Continuous (binned to 5 min)	Mix (direct average)	N/A
Mei et al. 2018 (19)	Per2-Venus; Cry1-Venus	SCN SPZ CA1/2	15 s/10 min	Mix (direct average)	No
Jones et al. 2018 (36)	GCaMP6s	SCN	10 min/60 min	Integration Peaks	Yes
Jones et al. 2021 (40)	Per2-Venus; GCaMP6s	PVN	60 s/15 min (Per2); 10 min/60 min (Calcium)	Mix (Per2) Integration (Calcium) Peaks (Calcium)	Yes
Maejima et al. 2021 (24)	GCaMP7s	SCN	30 s/10 min	Mix (direct average)	No
Zhai et al. 2022 (37)	GCaMP7s	SCN	30 s/10 min	Frequency	Yes
Xie et al. 2023 (68)	GCaMP6s	SCN	30 s/10 mina	Mix (direct average)	Yes
Tang et al. 2023 (47)	GCaMP7s	DMH	2 min/12 min; 10 min/60 min	Mix (direct average) Integration	Yes
Tsuno et al. 2023 (25 )	GCaMP7s	SCN	30 s/10 min	Mix (direct average)	No
Onodera et al. 2023 (69)	GCaMP7s	SCN	30 s/10 min	Mix (direct average)	No
Sayar-Atasoy et al. 2023 (38)	GCaMP7s	ARC DMH SCN	5 min/30 min; 10 min/30 min	Integration	Yes
Vas et al. 2024 (46)	GCaMP6s	ARC	5 s/20 s; Continuous	Peaks	Yes
Douglass et al. 2024 (39)	GCaMP6s	ARC PVN DMH	Continuous	Mix (direct average)	Yes
Gomez et al. 2024 (70)	GCaMP6s	PVN	10 min/60 min	Integration Peaks	Yes
Peng et al. 2025 (71)	GCaMP7s	SCN	30 s/10 min	Mix (direct average)	Yes
M. Wang et al. 2025 (72)	GCaMP7 s; jRGECO1a	SCN	30 s/10 min	Mix (direct average)	Yes
Yurgel et al. 2025 (73)	GCaMP8s-soma	SCN	4 min/12 min	Tonic	Yes
El Cheikh Hussein et al. 2022 (48)	GCaMP6f (single cell)	SCN	1 min/30 min	Tonic Integration Peaks	N/A
Stowie et al. 2023 (56)	GCaMP7s (single cell)	SCN	5 min/3 h	Mix (direct average) Peaks	N/A

This table compiles studies that use genetically encoded reporters to measure circadian rhythms in neural activity or gene expression in freely behaving animals. It highlights the diversity of recording strategies (e.g. continuous vs. intermittent sampling), target brain regions, and analytical approaches used to quantify both slow baseline trends and fast transient events. Studies are listed chronologically and include both fluorescent (e.g. GCaMP) and bioluminescent (e.g. Per1-luc) reporters.

SCN, suprachiasmatic nucleus; OB, olfactory bulb; SPZ, subparaventricular zone; CA1/2, hippocampal subfields CA1 and CA2; PVN, paraventricular nucleus; DMH, dorsomedial hypothalamus; ARC, arcuate nucleus.

^aSession length inferred from analysis code; not explicitly reported in the manuscript.

Circadian rhythms provide an ideal case study for refining such long-term recording strategies. As endogenous, near-24 h cycles that regulate a wide range of physiological and behavioral processes, they represent one of the most well-characterized biological systems operating across extended timescales (1, 2). At the cellular level, circadian rhythms are generated by transcriptional–translational feedback loops that govern the expression of “clock genes” and their downstream targets (3–5). In mammals, these rhythms are coordinated by the suprachiasmatic nucleus (SCN), the central circadian pacemaker that synchronizes peripheral clocks throughout the body via photic input from the retina (6).

Understanding circadian function requires continuous measurement across molecular, cellular, and behavioral domains. Historically, this has relied on indirect behavioral or physiological readouts such as locomotor activity, body temperature, or hormone secretion. While these approaches have provided valuable insight into systems-level rhythms, they do not directly measure the underlying neuron circuit dynamics. To more directly understand how the brain generates and maintains circadian rhythms, researchers have turned to methods such as c-FOS immunohistochemistry and in vivo electrophysiology (7, 8). However, these techniques lack the temporal continuity, resolution, or cell-type specificity needed for long-term circuit-level analysis.

GEFIs overcome many of these limitations by enabling long-term, cell-type specific recordings in individual animals with high temporal resolution. This article uses circadian rhythms as a case study to examine the challenges and best practices of long-term GEFI imaging. Importantly, the strategies we outline here are broadly applicable to other neural systems that regulate complex, slowly evolving behaviors and physiological processes such as feeding, hormone secretion, and thermoregulation.

A focus on calcium imaging

Recent advances have expanded GEFIs to include reporters not only for calcium (9–12) but also for neurotransmitters (e.g. glutamate, GABA) (13, 14), neuromodulators (e.g. dopamine, opioids) (15–17), second messengers (e.g. cAMP) (18), and even circadian clock gene expression (19). Additionally, genetically encoded voltage indicators (GEVIs) provide direct measurements of membrane potential, offering a more immediate readout of neuronal excitability (20).

Among these tools, genetically encoded calcium indicators (GECIs) are by far the most widely used, with the green fluorescent variants known as GCaMPs serving as the standard tool for monitoring neuronal activity (9–12). In neurons, action potentials open voltage-gated calcium channels, triggering calcium influx that can be detected by calcium indicators. Because calcium signaling is integral to synaptic transmission and intracellular communication, calcium imaging provides a robust proxy for neuronal firing and circuit-level dynamics.

Given their widespread adoption and extensive optimization, GECIs will be the focus of this article. However, the analytical challenges and recommendations we outline, particularly those related to signal decomposition, also generalize to other optical reporters that report dynamic neural signals over extended timescales.

Limitations of short-term analysis strategies for long-term fiber photometry

In freely behaving animals, fiber photometry provides a scalable, stable method for long-term GECI recordings. Genetic targeting allows recordings from specific neural populations, and the mechanical simplicity of the setup enables stable, multiday recordings with minimal motion artifact. Moreover, unlike imaging methods that produce large, complex datasets, photometry generates compact, continuous time series that are ideal for long-term recordings across animals and experiments.

These features have thus made fiber photometry a widely adopted method for monitoring neural activity over extended timescales. However, the accessibility and scalability of the technique have greatly outpaced the development of analytical strategies capable of handling long-duration, continuous recordings. Most existing approaches were designed for short-term, event-triggered experiments and assume relatively stable baselines, isolated transients, and clear stimulus–response relationships. However, these assumptions break down over extended recording durations.

In reality, fluorescence signals produced by GECIs represent a convolution of multiple biological and technical processes (e.g. calcium influx, protein turnover, intracellular signaling, photobleaching) that interact over vastly different timescales and must be disentangled to support meaningful interpretation. Disentangling these overlapping influences requires analytical frameworks specifically tailored to long-term recordings. Traditional ΔF/F₀ (change in fluorescence over baseline fluorescence) normalization, while effective for short-term experiments, can obscure actual biological dynamics by conflating slow shifts in baseline fluorescence with true changes in neural activity.

A reproducible, interpretable alternative: tonic and phasic signal decomposition

After surveying the literature on long-term optical recordings, we identified four primary analytical approaches, each targeting distinct aspects of the GECI fluorescence signal (Table 1). Although terminology varies across studies, these methods can be unified under a conceptual framework that distinguishes tonic from phasic signal components. Tonic signals reflect slow, sustained calcium fluctuations that are associated with circadian gene expression, intracellular calcium stores, or other, nonspiking regulatory processes. Conversely, phasic signals represent faster transients, typically reflecting action potentials or synaptic input. In fiber photometry, such transients are most detectable when activity is synchronized across a large population of neurons.

Critically, tonic and phasic signals are not defined by amplitude thresholds or exclusionary cutoffs. Instead, they reflect independent features of calcium fluorescence dynamics (slow and sustained vs. fast and transient) that can be analyzed separately or in combination depending on the scientific question. Separating these signal components improves interpretability, allows for consistent comparisons across studies and brain regions, and reduces the risk of drawing misleading conclusions. While this tonic/phasic framework emerged from circadian neuroscience, it applies broadly to any long-term experiment using GECIs to monitor neural activity. In the following sections, we outline the challenges of interpreting long-term fluorescence signals, describe strategies for extracting tonic and phasic signals, and propose best practices for data acquisition, analysis, and reproducibility.

Foundations of GECI-based optical recording

Signal components in long-term recordings

As discussed above, interpreting GECI signals over hours or days requires disentangling multiple biological and technical influences that shape the recorded fluorescence signal over time. These include both genuine sources of neural activity and confounding factors that introduce drift, noise, or distortions (Fig. 1).

Fig. 1.

Coupling between tonic and phasic calcium signals varies across brain regions in long-term photometry recordings. A) Simulated calcium-dependent and isosbestic signals over 3 days, illustrating two representative recording scenarios. The top trace models a region with strong circadian rhythms in tonic calcium (e.g. SCN), while the bottom trace models a region with relatively stable tonic calcium (e.g. ARC). Numbered timepoints (1–4) correspond to magnified views in (B and C) Schematic decomposition of fluorescence signals at each timepoint. Signals are conceptually separated into four components: activity-driven calcium, baseline intracellular calcium, ligand-free fluorescence, and autofluorescence. This decomposition is conceptual, intended to illustrate major sources of the fluorescence signal, and is not derived from experimental data. In regions like the SCN (panels 1–2), both tonic baseline and phasic activity fluctuate over time, resulting in higher transients with an elevated baseline during the day and fewer transients with a lower baseline at night. In contrast, regions like the ARC (panels 3–4) show a similar circadian variation in phasic activity but maintain a stable tonic baseline. This distinction highlights the need to analyze tonic and phasic signal components as distinct but complementary features of long-term recordings.

Autofluorescence

Autofluorescence arises primarily from optical components and endogenous fluorophores in brain tissue, and gradually diminishes due to photobleaching over time (21). Although unavoidable, this background signal can obscure low-amplitude changes in fluorescence and must be carefully accounted for during data analysis.

Ligand-free baseline fluorescence

GECIs are engineered by linking calcium-binding domains to a fluorescent protein, typically a circularly permuted single fluorophore (9, 12, 22). Calcium binding induces a conformational change in the fluorescent protein domain, altering its fluorescence intensity. However, because these indicators are fluorescent proteins, they exhibit baseline fluorescence even in the absence of calcium. A key goal in GECI development has been to minimize this baseline fluorescence to improve contrast (11).

Ideally, baseline fluorescence would remain constant over time for a given number of fluorescent proteins. In practice, however, it is influenced by photobleaching and protein turnover. In long-term recordings, this becomes particularly important, as the endogenous circadian regulation of protein synthesis and degradation can introduce rhythmic fluctuations in baseline GECI levels (23–25).

Neuronal activity-driven calcium transients

Once baseline shifts are accounted for, the primary use of GECIs is to detect calcium transients linked to neuronal activity, including action potentials and synaptic activity (26). The most common readout in fluorescent calcium imaging is the relative change in fluorescence (ΔF/F₀). Various methods exist for defining baseline fluorescence (F₀), such as using prestimuli brightness levels or calculating a moving average in a sliding time window (12, 27, 28). For short-term recordings (seconds to minutes) in response to discrete stimuli, ΔF/F₀ is an effective, versatile measure of neuron activity, as transient calcium influx is closely tied to neuronal firing and synaptic transmission (26, 29, 30).

However, when applied to long-term recordings over hours to days, the interpretation of both calcium transients and ΔF/F₀ itself becomes more complex. In these cases, spontaneous activity must be carefully analyzed, as calcium transients can be influenced by underlying mechanisms beyond somatic action potentials (31).

Baseline intracellular calcium

Acute calcium imaging studies often do not emphasize baseline calcium levels (F₀) when normalizing results to ΔF/F₀, as their primary focus is on transient fluorescence changes following experimental stimuli. However, intracellular calcium serves as a critical second messenger and is subject to regulation beyond neuronal spiking activity. Due to dynamic changes in channel conductance, ion buffering, and release from internal stores, baseline intracellular calcium levels can fluctuate independently of action potentials. These fluctuations typically occur more slowly and tend to stabilize at an equilibrium level over short time windows (i.e. several minutes) (26). For example, in the SCN, intracellular calcium levels exhibit circadian fluctuations on a much longer timescale than action potential-driven calcium transients. Notably, these slow calcium rhythms do not fully align with the circadian pattern of spontaneous neuronal firing (32–34). Intracellular calcium fluctuations in the SCN, and most likely, in other brain regions, therefore represent an independent regulatory process, not merely noise.

The role of internal controls

Long-term GECI recordings are particularly susceptible to slow, cumulative signal artifacts, including autofluorescence, photobleaching, animal and fiber motion, and circadian fluctuations in protein turnover. To minimize these effects, proper use of control signals is essential for data interpretation.

The most commonly used internal control in fiber photometry is the isosbestic signal, which is generated by exciting the GECI at a wavelength where its emission is independent of calcium binding. Because this signal originates from the same cells and fluorophores as the calcium-sensitive channel, it acts as a valuable reference for correcting calcium-independent fluorescence artifacts. As an alternative (for instance, when experimental hardware does not support isosbestic excitation), researchers can co-express a stable, ligand-insensitive fluorophore (e.g. GFP, tdTomato) alongside a GECI of a different color in the same animal (35). Both isosbestic signals and signals from co-expressed stable fluorophores can be processed similarly.

However, internal control signals have limitations. Because the isosbestic point does not correspond to the optimal excitation wavelength for most fluorophores, the emitted isosbestic signal is often weaker than the ligand-sensitive channel. Moreover, many commercially available fiber photometry systems use fixed wavelength “isosbestic” excitation (e.g. 405 nm) that deviates slightly from the true isosbestic point for many GECIs, leading to partial calcium sensitivity in the “control” channel. In some cases, this can even yield an inverted calcium response. The ability of the isosbestic signal to track artifacts will also degrade over time due to photobleaching during extended recordings.

To mitigate these risks, we recommend including external controls: animals expressing only a stable, ligand-insensitive fluorophore in the same brain region and cell type, using the same viral vector and promoter. These controls provide a stable reference for detecting artifactual changes in fluorescence that may otherwise be interpreted as neuronal activity. The most ideal external control is to use a ligand-insensitive mutant version of the fluorescent sensor itself (15), though this is not typically available.

Notably, even “stable” fluorophores are not immune to changes over time. Small, but detectable, circadian variations in EGFP fluorescence have been observed in the SCN (24, 25). These fluctuations appear to affect only the tonic component of the fluorescence signal and differ in both amplitude and phase from the rhythms observed in GECI-expressing animals (24, 25, 36), possibly reflecting the circadian regulation of the promoter driving fluorophore expression or protein turnover. These findings highlight that even stable, ligand-insensitive fluorophores may exhibit weak rhythmicity under certain conditions. Thus, including well-matched external controls is critical for not only identifying artifactual rhythms but also for correctly interpreting the origin and nature of fluorescence fluctuations in long-term imaging data.

Signal decomposition for long-term recordings

We previously introduced the distinction between the tonic and phasic signal components as a conceptual tool for interpreting long-term calcium imaging data. Importantly, these components are not mutually exclusive subsets of the signal, but rather represent distinct analytical modes that extract complementary information from the same fluorescence trace. Although few studies use these specific terms, most analytical approaches in the literature target one or the other of these components in practice without explicitly defining them. For instance, some approaches prioritize sustained shifts in baseline calcium to capture slow intracellular dynamics (what we define as “tonic”) (Figs. 2 and 4A), while others isolate discrete calcium transients as a proxy for neuronal firing (what we define as “phasic”) (Figs. 3 and 4A). Here, we expand upon this conceptual framework to outline analytical strategies designed to extract tonic and phasic signals more reliably and reproducibly across long-term recordings.

Fig. 2.

Workflow for extracting tonic calcium signals from long-term photometry recordings. A) Raw calcium-dependent and isosbestic signals recorded from GCaMP7s-expressing SCN NMS neurons over 2 days. Data were acquired in 10 min sessions once per hour; 50 min “off-periods” are omitted for illustration. Inset: magnified view of a single 10 min session. B) The isosbestic trace is “globally fit” across the full 48 h recording to correct for photobleaching, changes in protein expression, and low-frequency motion artifacts. C) Within each session, ΔF/F₀ is computed as (signal—fitted isosbestic)/fitted isosbestic, using the portion of the global isosbestic fit corresponding to that session's time window. The tonic signal is then defined as the 2nd percentile of ΔF/F₀ values within each 10 min session. D) Tonic calcium and average fluorescence signals over time. While both signals show circadian variation, because the average fluorescence trace includes both tonic and phasic signal components, its peak time occurs later than that of the percentile-defined tonic signal. Inset: magnified view of the phase difference (triangles).

Fig. 4.

Signal decomposition framework and recommended practices for long-term photometry analysis. A) Conceptual pipeline for separating raw fluorescence signals into tonic and phasic signal components. Each branch highlights distinct steps for artifact correction, ΔF/F₀ calculation, and signal quantification, reflecting their different temporal dynamics and analysis needs. B) Recommended practices for acquisition, analysis, and visualization of long-term recordings. These guidelines support reproducibility and ensure proper interpretation of signal features in datasets recorded over extended timescales.

Fig. 3.

Workflow for extracting phasic calcium signals from long-term photometry recordings. A) Example 10-min session from the SCN recording shown in Fig. 2 illustrating phasic signal extraction following artifact correction. Left: the calcium-dependent signal and isosbestic reference share an “upward” artifact. Center: least-squares (LSQ) fitting, the most widely used method for isosbestic correction in fiber photometry, often fails to capture such transient artifacts, flattening the fitted isosbestic trace. Alternatives such as correlation-guided “dynamic” fitting adapt to local signal structure and preserve shared artifacts in the fitted isosbestic. Right: As a result, LSQ correction introduces a spurious upward deflection in the ΔF/F₀ trace, whereas dynamic fitting avoids this distortion. B) Stacked 10 min ΔF/F₀ traces recorded hourly across a full circadian cycle. ZT (zeitgeber time) indicates hours since lights-on in a 12h:12 h light:dark cycle, where ZT 0 is lights-on and ZT 12 is lights-off. C) Representative 10 min ΔF/F₀ traces depicting potential patterns of phasic activity. The shading indicates signal used to calculate integrated calcium levels; the triangles indicate detected peaks classified as calcium transients. While the top and bottom traces have similar peak counts, they differ substantially in integrated calcium levels due to differences in transient amplitude and shape. This divergence can occur in both directions: sessions with similar numbers of transients may yield different integrated calcium levels, and sessions with similar integrated calcium levels may differ in peak frequency. For accurate characterization, both metrics should be reported when feasible. D, E) Normalized calcium transient numbers (D) and integrated calcium fluorescence levels (E) measured from each 10 min ΔF/F₀ trace across 48 h. Note that both measures reveal circadian variations in phasic activity, but with distinct temporal profiles.

Analytical strategies for tonic signal extraction

The pitfalls of sessional averaging

To minimize phototoxicity and photobleaching, many long-term optical imaging studies acquire data in discrete sessions (e.g. 10 min intervals per hour; Table 1). A widely used analysis method is to simply average the fluorescence signal within each session. However, because these sessions often span several minutes, this approach obscures fast, transient calcium events and conflates them with slower changes in baseline fluorescence (Fig. 2D).

The consequences of this averaging can vary by brain region. In the SCN, baseline intracellular calcium fluctuations exhibit a greater amplitude than action potential-driven calcium transients (37). Consequently, averaging fluorescence predominantly reflects the circadian rhythm of baseline intracellular calcium, rather than neuronal firing patterns. In contrast, brain regions such as the arcuate nucleus (ARC) and the paraventricular nucleus of the hypothalamus (PVN), exhibit circadian rhythms in fast calcium transients that are often more pronounced than baseline calcium levels (38–40). As a result, sessional averaging does not represent calcium dynamics uniformly across brain regions, as the relative contributions of baseline calcium and transient calcium signals vary depending on the neural circuit being studied.

To overcome this limitation, tonic activity should be extracted from continuous recordings using analysis methods designed to preserve genuine fluctuations in intracellular calcium while minimizing technical artifacts such as motion, photobleaching, and protein turnover.

Global isosbestic fitting

A robust strategy for extracting tonic dynamics from long-term GECI recordings is to fit the calcium-dependent signal to an internal control signal (typically the isosbestic channel) across the entire recording period, treating all sessions as a continuous time series (Fig. 2B). This “global isosbestic fitting” approach minimizes session-to-session variability while preserving slower drifts in baseline fluorescence over time.

To implement this, we perform a linear least-squares fit between the isosbestic and calcium-dependent signals across the entire dataset. This global fit effectively centers the isosbestic signal within the calcium-dependent signal, allowing for the optimal correction of signal artifacts while preserving the underlying tonic circadian oscillation.

However, the relationship between the isosbestic and calcium-dependent signals is not always stationary and may include a nonlinear drift over time. As a result, residual slow trends may persist even after global fitting. To correct for this, a 24 h sliding average can be subtracted from the trace, following approaches adapted from long-term circadian bioluminescence recordings (41, 42).

Defining tonic baselines

A key methodological challenge in tonic signal extraction is distinguishing tonic signals from phasic transients that “ride on top of” the tonic signal. Some studies attempt to isolate the tonic baseline by refitting the isosbestic signal within each recording session. However, this strategy can inadvertently absorb phasic transients into the tonic baseline, especially during periods of heightened neuronal activity, when large or overlapping calcium transients distort the local isosbestic fit. This is further complicated because many commercial systems use excitation wavelengths that do not precisely match the isosbestic point of the fluorophore, introducing residual calcium sensitivity into the isosbestic control signal.

To mitigate these risks, we recommend defining a fixed threshold across the full dataset, such as the 2nd percentile of the calcium-dependent fluorescence intensity distribution of each session, as the tonic baseline (Fig. 2C). This percentile-based thresholding approach provides a conservative, stable baseline that is resistant to fluctuations in phasic activity and ensures that normalization is applied uniformly across different experimental conditions, reducing biases introduced by session-to-session variability. Importantly, this threshold serves as an anchor for estimating slow, sustained shifts in intracellular calcium, not as a strict cutoff between tonic and phasic activity.

Analytical strategies for phasic signal extraction

Correcting artifacts in phasic signal extraction

Extracting meaningful phasic activity from long-term fiber photometry recordings requires careful separation of true calcium-dependent transients from slowly evolving artifacts such as motion, photobleaching, and baseline drift. These confounds accumulate over time and can obscure low-amplitude signals, particularly when neural activity is sparse or intermittent. Importantly, in contrast to short-timescale recordings where artifacts may remain relatively stable or negligible, long-term recordings demand artifact correction strategies that can adapt to changes in signal structure across minutes, hours, and days.

A common strategy for artifact correction uses the isosbestic signal as a reference to estimate and subtract motion- and photobleaching-related noise from the calcium-dependent signal (21). This typically involves fitting the isosbestic signal to the calcium-sensitive signal and removing the fitted contribution as a proxy for shared artifacts. However, the accuracy of this correction depends heavily on the choice of fitting strategy: different approaches vary in how well they adapt to changing artifact structure and preserve real neural transients, especially over long timescales (Fig. 3A).

The most straightforward approach to artifact correction is ordinary least-squares regression, a linear fitting model between the calcium-dependent and isosbestic signals (Fig. 3A, top). This method assumes a stable relationship between channels over time and can be effective for short-term or low-noise recordings. More robust regression methods, such as iteratively reweighted least squares, modestly improve fitting by reducing the influence of outliers and increasing fitting stability (43). However, in long-term recordings, the relationship between the signal and isosbestic channels often changes over time due to changes in motion, photobleaching rates, or recording stability. Static fittings models cannot account for these dynamics, which can lead to overcorrection in some portions of the recording and undercorrection in others.

To address the limitations of least-squares fittings, our lab has implemented a time-localized or “dynamic” isosbestic fitting strategy (Fig. 3A, bottom). In this approach, the recording is divided into short, overlapping windows (e.g. ∼1 minute), and a separate linear regression is performed in each. Heuristic rules, such as correlation-dependent scaling of the isosbestic amplitude, are then used to adapt the fit based on signal similarity within each window. For instance, when the Pearson correlation between the two channels is high, the isosbestic trace is scaled to more closely match the calcium signal under the assumption that noise dominates the recording. Conversely, when the correlation is low, the scaling factor is reduced to avoid subtracting real calcium transients that are absent in the isosbestic channel.

While these dynamic fitting strategies offer greater flexibility than linear regression, their performance depends on assumptions about the relationship between artifacts and real transients that may not hold across all brain regions, behavioral states, or recording conditions. To overcome these limitations, probabilistic models offer a more principled foundation for artifact correction. Our lab has recently developed a Bayesian regression framework that treats the isosbestic signal as a latent variable shaped by motion and photobleaching processes. This approach estimates the most likely artifact at each timepoint while simultaneously quantifying the uncertainty in the fit. Although still under active development, such models represent a promising direction for extracting phasic signals from long, artifact-prone recordings where deterministic fits fall short.

Separating correction and normalization

A common point of failure in ΔF/F₀ analysis is using the same fitted signal for both artifact correction and baseline normalization. This becomes especially problematic in brain regions like the SCN, where tonic calcium levels vary strongly across the circadian cycle. In these contexts, the raw isosbestic signal remains relatively stable but the calcium-dependent signal rises and falls with intracellular calcium levels. If the isosbestic trace is scaled to match the calcium-dependent signal within each session, the fitted isosbestic signal is elevated during times of high tonic calcium (e.g. daytime) and suppressed during times of low tonic calcium (e.g. nighttime). This causes phasic transients to be normalized to different baselines, making daytime transients appear artificially smaller and nighttime transients appear artificially larger even if their absolute amplitudes are identical.

To avoid this distortion, artifact correction (the numerator, or ΔF) and baseline normalization (the denominator, or F₀) should be explicitly separated. Specifically, the fitted isosbestic signal, scaled to capture slow, shared artifacts, should be used solely for subtraction during artifact correction. Baseline normalization, in contrast, should rely on an independent reference, ideally one that is both calcium-insensitive and stable across time such as the raw isosbestic trace from a given recording session or a low-percentile value from the isosbestic signal fitted over the full recording. Overall, this approach ensures both locally effective artifact correction and globally consistent normalization, preserving true phasic dynamics regardless of the underlying tonic signal.

Detecting phasic events: peak counting

Once artifacts have been minimized, the next step is to identify discrete calcium transients (Fig. 3C and D). To do this, many studies have used event detection algorithms to count the number of fluorescence peaks exceeding a predefined threshold. This strategy, which mimics firing rate analyses commonly used in electrophysiology, has been used to identify circadian rhythms in neural activity in several brain areas including the SCN, PVN, and ARC (Table 1).

However, it is important to reiterate that changes in GECI fluorescence do not directly correspond to action potentials. GEVIs provide a more direct readout of spiking, whereas GECIs respond more slowly and integrate calcium influx over time. Moreover, the temporal profile and “shape” of fluorescence transients can vary significantly across brain regions and behavioral states due to both the slower kinetics of GECIs and the population-level recording inherent to fiber photometry (28, 36, 38, 44–46).

Peak detection algorithms typically require user-defined parameters such as minimum peak height, width, and prominence. These parameters need to be optimized for specific cell types, circuits, and behaviors, and may not generalize across experimental contexts. Researchers applying peak detection in new systems or circuits should carefully validate detection settings to avoid misclassification or missed phasic calcium events. Even with optimized detection parameters, however, peak counting alone provides an incomplete picture of neuronal activity as peak counts do not capture the shape or magnitude of phasic transients. For example, a signal with sparse, low-amplitude peaks may yield a similar peak count to one with infrequent, but high-amplitude, synchronized bursts of activity, despite reflecting fundamentally different neural activity patterns.

Detecting phasic events: integrated fluorescence

To complement peak count analyses, many studies also report integrated calcium fluorescence as a way to capture overall phasic neural activity over time (Fig. 3C and E). This approach aims to provide a more holistic measure of neural dynamics by quantifying the cumulative fluorescence signal associated with calcium transients. In this approach, which we consider as a complementary measure of the phasic signal, a cutoff line defined within each recording session is used to isolate calcium transients from tonic baseline activity. Signal values above this threshold are quantified as phasic activity for the purposes of integration. Values below this reference are not assumed to be purely tonic (as they may include low-amplitude phasic activity or biological drift) but are excluded from this specific analysis step to reduce noise. Integrated phasic activity is then quantified by calculating either the integrated area between the fluorescence trace and this cutoff line (36, 47), or, in some cases, the average fluorescence value above this threshold (38). Different studies have used various strategies to define this cutoff, including the 10th percentile of the session (36); the 10th percentile of a sliding 30-s window (47); or a fitted isosbestic reference signal (38, 47). A conceptually similar approach has also been applied to single-cell SCN recordings, where researchers calculated the amplitude between the 5th to 95th percentile of the ΔF/F₀ as a measure of calcium transient variability (48).

As with peak counting, however, relying on fluorescence integration alone can obscure important details about the structure of individual calcium transients. For example, a recording session with a few large calcium transients may yield a similar integrated value to one with many smaller transients despite reflecting distinct neural dynamics. To more accurately capture phasic activity, peak count and integrated fluorescence should be reported together whenever possible (Fig. 3C).

Detecting phasic events: frequency-based decomposition

Finally, frequency analysis methods such as fast Fourier transforms or wavelet analysis can reveal dynamic changes in neural activity not captured by peak counting or integration. These methods preserve information about both the duration and magnitude of transient fluorescence changes, allowing researchers to identify changes in multiple oscillatory features of neural activity over time. By generating spectrograms that represent power across frequency bands, frequency analysis can easily distinguish between recordings with few high-amplitude transients, many low-amplitude transients, or sparse activity, overcoming key limitations of both peak finding and fluorescence integration.

So far, this analysis method has only been applied in one long-term fiber photometry study, which reported that Vgat1-positive SCN neurons exhibit greater high frequency spectral power during the daytime, suggesting a time-of-day dependent modulation of fast calcium dynamics (37).

Best practices for experimental design and data acquisition

Tradeoffs between commutators and untethered patch cords

Most commercial fiber photometry systems are wired, requiring implanted animals to be tethered to a fiber optic patch cord. In standard fiber photometry experiments, a commutator (also known as a rotary joint or slip ring) is commonly used to allow free movement and to relieve torque when animals rotate. However, commutator use presents a tradeoff: while it improves mechanical stability, it also introduces significant signal damping, which reduces signal amplitude and lowers the signal-to-noise ratio.

Signal damping becomes more pronounced in long-term recordings, where cumulative photobleaching further reduces signal strength. In brain regions such as the SCN, where phasic calcium transients are of relatively low amplitude, this degradation can severely limit signal detectability. In contrast, larger tonic oscillations in baseline intracellular calcium often remain detectable under these conditions. Consequently, commutator-based systems may be suitable for capturing slow, high-amplitude tonic rhythms, but less effective at studying dynamic, but low-amplitude, phasic activity.

To preserve signal fidelity in long-term recordings, some researchers choose to omit using a commutator and instead use lightweight, untethered fiber optic patch cords. While this avoids the signal damping introduced by the commutator, it increases the risk of motion artifacts from cord tension or tangling. To mitigate these issues, the patch cord should be anchored well above the cage floor (typically around 40 cm or higher) to allow freedom of movement and reduce strain at the connection to the animal. Any resulting motion artifacts should be corrected using internal control signals and robust analytical methods that can reliably extract meaningful signals despite the challenges of prolonged recordings.

Wireless recording devices offer a promising alternative to tether-related artifacts, allowing for greater freedom of movement and supporting more complex behavioral paradigms. However, current systems are constrained by limited battery life, device size, and data transmission stability. Significant technical improvements are still needed before wireless photometry becomes a practical and widely adopted option for high-resolution, multiday recordings.

Optimizing mechanical coupling between the patch cord and implant

A critical source of signal instability is the physical connection between the implant and patch cord. The most common connector is a ceramic or metal mating sleeve, which relies on friction to hold the ferrules together. However, this interface is susceptible to micromovements that can accumulate or drift over time during long-term recordings. To minimize these artifacts, researchers often apply a thin layer of dental cement or nail polish around the sleeve to stabilize the connection. A more robust, though less widely adopted, alternative is a quick-release interconnect, which physically locks the patch cord to the implant using a mechanical coupling instead of friction. This fixed connection greatly improves mechanical stability, making it a strong choice for experiments requiring consistent signal transmission over time.

Minimizing phototoxicity and photobleaching

Long-term fluorescence recording requires careful consideration of phototoxicity (Fig. 4B). Shorter excitation wavelengths, such as the ∼470 nm blue light used for GCaMP, are more phototoxic to live cells (49) and have been shown to reduce circadian rhythmicity in cultured SCN slices compared to red light (50). Although fiber photometry uses light intensities that are typically 2–3 orders of magnitude lower than those used in optogenetics, phototoxicity remains a concern. One long-term circadian fiber photometry study used a TUNEL assay to assess phototoxicity and found no evidence for phototoxicity-induced apoptosis when using a 15 µW laser active for 15 s every 10 min (19). A similar assay is recommended when using higher laser or LED power, longer recording sessions, or extended total recording durations.

Minimizing fluorescence photobleaching is also essential for extended recordings. To address this, published studies (Table 1) typically use session-based recording schedules rather than continuous illumination. Additionally, excitation light power may need to be reduced compared to levels used for acute recordings. To ensure accurate estimates of tissue-level light exposure, researchers should measure and report light power at the tip of a fiber optic cannula connected to the recording setup, as this provides a more meaningful estimate than measurements taken directly from the patch cord alone.

Managing crosstalk and ambient light

When recording both isosbestic and calcium-dependent signals, which share an emission wavelength but have different excitation wavelengths, proper signal separation is essential. While we do not endorse a specific approach, we highlight key considerations for minimizing artifacts during both acquisition and analysis.

In method 1, two excitation lights are alternately triggered, creating separate signal channels indexed by excitation wavelength. However, any light captured during each excitation pulse, including ambient light, can contaminate the signal. To prevent this, the connection between the patch cord and the implanted cannula should be light-tight. This can be achieved by using mating sleeves covered with black heat-shrink tubing or other opaque interconnects. Additionally, covering the dental cement cap around the fiber optic implant with black nail polish helps prevent ambient light from reaching the fiber through the skull, which is particularly important in brightly lit behavioral setups.

In method 2, each excitation light is modulated at a distinct frequency, allowing a lock-in amplifier to demodulate the shared emission signal and separate the contributions from each excitation channel. Since ambient lighting sources fluctuate at known frequencies (e.g. 100–120 Hz for fluorescent lights, and several kHz for LEDs) (51), modulation frequencies should be chosen to avoid both these frequencies and their harmonics, that is, integer multiples of the base frequency. This frequency-based method effectively eliminates ambient light contamination as unmodulated environmental light does not contribute to the demodulated signal. However, it requires continuous illumination from both excitation sources, increasing the risk of photobleaching and phototoxicity during long recordings.

Session scheduling: duration vs. frequency tradeoffs

A common strategy to minimize photobleaching and phototoxicity is to collect data in discrete recording sessions rather than continuously (Fig. 4B). While the overall duty cycle remains constant, researchers can tune temporal resolution by adjusting session length and frequency. For example, recording for 10 min every hour or 2 min every 12 min results in the same total light exposure time (10 min per hour), but the latter samples more frequently and thus provides finer temporal resolution.

The optimal session length depends on the signal of interest. For tonic signal analysis, which typically reduces each session to a single value (e.g. the second percentile of the ΔF/F₀), the fine-grained waveform structure within each session is less critical. In these cases, shorter and more frequent sessions (e.g. 30 s every 10 min) can enhance temporal resolution across the recording period without compromising interpretability. In contrast, when the primary goal is to extract phasic calcium transients associated with neural activity, longer recording sessions (e.g. 10 min every 60 min) are often preferable. These provide sufficient data for robust fitting of the isosbestic signal and more reliable separation of “true” calcium events from noise. Importantly, because many long-term recordings contain both slow tonic fluctuations and fast phasic transients, session design should be tailored to preserve the temporal resolution required for each signal component based on specific experimental and analytical goals.

Tools and recommendations for reproducible analysis

Open-source analysis toolkits: strengths and limitations

User-friendly open-source packages such as GuPPy, pMAT, and PyFiber have substantially lowered the barrier to entry for fiber photometry analysis (52–54). These tools simplify common preprocessing steps and event-triggered analyses, enabling researchers with limited programming experience to extract meaningful results from short-term experiments.

However, most of these tools were developed with acute or subacute experiments in mind. Their design assumptions (short recordings, single-session file structures, time-locked behavioral events, etc.) limit their applicability to long-term studies involving spontaneous behavior or circadian timescales. As the field shifts toward chronic monitoring of neural dynamics, this creates a significant bottleneck in analysis standardization.

Core requirements for long-term photometry analysis

Long-term fiber photometry datasets often consist of hundreds or thousands of session files, or continuous recordings spanning multiple days. These datasets exceed the capacity of tools designed for manual file-by-file processing. Efficient preprocessing and quality control of these large datasets requires automated, scalable pipelines, which are currently absent from most general-use analysis packages.

Equally problematic is the heavy reliance on stimulus-locked analysis frameworks, such as peri-event time histograms or trial-based averaging. These methods are appropriate for detecting stimulus-evoked transients but poorly suited to circadian or spontaneous paradigms, where no discrete triggering event defines the structure of the data. Analyses in these contexts must instead focus on extracting slow, rhythmic tonic dynamics, capturing phasic signal patterns in the absence of external triggers, and characterizing within-subject changes across hours or days.

Most existing pipelines also treat slow fluctuations in fluorescence as baseline drift to be removed. This assumption undermines the analysis of tonic components, which often encode meaningful biological rhythms. Without explicit decomposition of tonic and phasic signals, long-term recordings risk being misinterpreted or oversimplified.

Another critical requirement is the accurate modeling of the isosbestic control signal. Over long timescales, even small fitting errors can accumulate, distorting ΔF/F₀ signals and introducing artifacts that confound interpretation. Robust and flexible control fitting strategies are essential for long-term studies but remain underdeveloped in current toolkits.

Modularity and scalability

To address these limitations, future photometry analysis platforms must prioritize scalability, modularity, and standardization. Scalability means enabling automated batch processing of large datasets with minimal manual intervention; modularity refers to the ability to plug in new components (e.g. preprocessing routines, analysis types, signal decomposition modules) without rewriting the full analysis pipeline; and standardization ensures that results can be reproduced across labs and datasets, with shared conventions for input formats, output structures, and signal processing steps.

Cloud-compatible pipelines or high-performance computing support can alleviate bottlenecks in data throughput and encourage broader data sharing. Tools that meet these criteria will allow long-term fiber photometry to move from a custom-coded, lab-specific approach to a generalizable and trusted methodology for analyzing neural activity across timescales.

Recommended practices for visualizing long-term recordings

In addition to a well-structured analytical pipeline, long-term photometry analysis must include clear, interpretable data visualizations that reflect both the structure and variability of the underlying signal, along with a detailed description of the data processing workflow. This practice allows for a clearer evaluation of analysis choices and facilitates comparison of results across studies. We recommend researchers include the following data visualizations (Fig. 4B):

1. Representative, single-session raw traces (Fig. 3A) of both calcium-dependent and isosbestic signals prior to normalization or artifact correction. This allows reviewers and readers to evaluate preprocessing choices, signal quality, and the impact of artifact correction.

2. Continuous trace overlays (Fig. 2A) that display the calcium and isosbestic signals over an extended timescale (e.g. several days) to highlight broad circadian patterns (38, 39, 47). Visual inspection of these traces helps determine whether a recording is dominated by tonic or phasic activity, an essential first step for selecting the appropriate analysis strategy. For example, in the SCN, tonic calcium exhibits a strong sinusoidal rhythm, while phasic calcium exhibits a square-wave-like pattern with a moderate amplitude (25, 36, 37). In contrast, the ARC and PVN show stronger phasic rhythmicity, with tonic and phasic signals peaking at different circadian times (38–40).

3. Stacked session-aligned traces (Fig. 3B) plot ΔF/F₀ waveforms from each recording interval across the circadian cycle (e.g. ZT or CT 0-23) using consistent scaling (36, 38, 40). This stacked panel of traces over time is especially useful for visualizing phasic signals, allowing for a direct assessment of waveform shape, signal variability, and the presence of transients over time that are often obscured by summary statistics or in continuous trace overlay plots alone.

Presenting raw traces, continuous, overlays and session-aligned stacks allows researchers to evaluate the structure and timing of both the tonic and phasic components of the fluorescence signal with greater clarity. These practices promote transparency, support informed analysis decisions, and help ensure that biologically meaningful circadian dynamics are not lost during data processing.

Emerging innovations and broader applications

Extending the tonic/phasic framework beyond GECIs and conventional photometry

As mentioned above, recent advances have expanded the GEFI toolkit beyond calcium to include sensors for neurotransmitters, neuromodulators, second messengers, and even circadian gene expression. Many of these indicators report signals that exhibit both tonic (slow) and phasic (fast) dynamics, much like GCaMP. Indeed, the terminology used throughout this article is itself inspired by the well-established framework of tonic and phasic dopamine signaling (55). Similarly, voltage signals captured by GEVIs reflect both “tonic” subthreshold membrane fluctuations and “phasic,” discrete action potentials. These parallels reinforce the broader applicability of the recording and analysis strategies discussed here to any long-term optical recordings targeting dynamic neural signals.

To achieve both cellular resolution and long-term recording capability, miniature fluorescence microscopes (“miniscopes”) have been developed to enable single-cell imaging in freely moving animals. By implanting a gradient refractive index lens above brain structures such as the SCN, researchers can track activity of individual neurons across multiple circadian cycles (48, 56). This approach provides cellular-level spatial resolution and detailed visualization of network dynamics that are inaccessible to bulk recording methods. Continued advances in hardware and analytical pipelines (57) have made miniscope systems increasingly accessible for chronic recordings. Long-term miniscope imaging has also been extended to other domains such as learning and memory (58), where repeated recordings across several hours captured circuit dynamics over extended timescales. Although so far these paradigms span shorter overall durations than multiday circadian recordings, they exemplify how cellular-resolution imaging can probe circuit dynamics over behaviorally relevant timescales. Importantly, analyses of miniscope datasets reveal both tonic and phasic calcium dynamics (48, 56), suggesting that applying the conceptual framework developed here could uncover additional insights from sustained tonic fluctuations often overlooked by conventional analyses focused solely on transient activity.

Another emerging advance of optical imaging technique is fluorescence lifetime photometry. This technique measures the decay time of fluorescence rather than its intensity, making it less susceptible to artifacts from photobleaching or fluctuations in excitation light. This improves signal stability in long-term recordings and allows signals to be reported in absolute rather than relative units, simplifying normalization and supporting cross-experiment comparisons. While specialized equipment and compatible sensors are required, recent technical advances are making the approach more accessible. Importantly, fluorescence lifetime still reflects underlying molecular activity, so distinguishing tonic and phasic components remains just as critical as in conventional photometry.

Low duty-cycle continuous photometry

Long-term fiber photometry poses a unique challenge in balancing prolonged data collection with the risks of photobleaching and phototoxicity. Most alternating light pulse paradigms operate at a 50% duty cycle (meaning that the light is on half the time), which typically restricts recordings to short sessions. Recently, Douglass and colleagues overcame this limitation by recording from ARC and PVN neurons using 0.75 ms excitation pulses at 20 Hz, resulting in a total duty cycle of just 1.5% (39). In this design, the excitation light is only on for a fraction of a second, but the pulses themselves occur every 50 ms. Because GCaMP fluorescence decays slowly relative to this sampling rate, the individual fluorescence events effectively “blend together” into a continuous signal. This approach therefore enables uninterrupted, high-resolution recordings while drastically reducing total light exposure.

While this recording strategy offers an advantage over session-based recording by enabling continuous, high-resolution data acquisition with minimal light exposure, the subsequent analysis did not distinguish between tonic and phasic calcium signals. Instead, the data were smoothed using a 30 min moving window, which effectively blended the signals and contributed to a roughly 3 to 4 h difference in the reported peak phase of calcium activity in the ARC and PVN compared to studies that specifically analyzed phasic activity (38–40). These discrepancies highlight the importance of separating tonic and phasic signals during analysis to ensure the accurate interpretation of long-term calcium dynamics.

Applications beyond circadian biology

Although this article focuses on circadian recordings, the logic of tonic/phasic decomposition generalizes to other behaviors and physiological processes that unfold over hours or days. Long-term fiber photometry has recently been applied to investigate the neural dynamics underlying feeding and drinking (59), thermoregulation (60), torpor (61), the estrous cycle (46), lactation (62), and epilepsy (63). These studies span recording durations from tens of minutes to several hours, with only one noncircadian study to date extending into a true multiday timescale (46). However, in most cases, these studies rely on analysis pipelines developed for short-term experiments and fail to distinguish between the tonic and phasic components of neural activity.

This distinction is especially important because regardless of the circuit or behavior being investigated, calcium fluorescence inherently reflects both slow, sustained intracellular calcium changes (driven by second messenger signaling) and rapid transients (as a proxy for neural activity). These components are often visible even over short timescales. For example, food delivery rapidly suppresses both the baseline calcium signal and frequency of “superimposed” transients in agouti-related peptide neurons in the ARC (44). In contrast, during torpor, neurons in the medial and lateral preoptic area exhibit a decrease in baseline calcium fluorescence but an increase in transient calcium events (61). Although these dynamics occur over minutes or hours rather than days, they reflect the same underlying principle as the tonic and phasic components we describe for circadian recordings. While not explicitly categorized in this way, the distinct temporal features of these calcium signals suggest that separating tonic and phasic signal components for analysis could reveal previously unappreciated patterns in long-term neural activity.

Conclusions

Maximizing the utility of long-term optical recording requires more than just technical capability. It demands clear experimental design, transparent data processing, and consistent analytical practices. This article addresses these needs by outlining best practices for acquiring and analyzing long-term GEFI recordings, with a particular emphasis on distinguishing the tonic and phasic components of the fluorescence signal. Recognizing these signal components as complementary, rather than competing, modes of analysis allows for richer interpretations of long-term neural dynamics, especially when studied in parallel. Looking ahead, progress will depend on scalable, open-source analysis tools, standardized reporting conventions, and collaboration across labs to ensure that results are both reproducible and broadly interpretable. These efforts will ultimately accelerate discovery of how neural activity evolves over long timescales across diverse circuits, behaviors, and physiological processes.

Funding

This work was supported by funding National Institutes of Health (NIH) 1ZIAMH002964 (S.H.), National Institutes of Health (NIH) R35 GM151020, and the Whitehall Foundation (J.R.J.).

Author Contributions

Qijun Tang (Conceptualization, Formal analysis, Supervision, Methodology, Writing—original draft, Project administration, Writing—review & editing), Jordan N. Cook (Conceptualization, Formal analysis, Supervision, Methodology, Writing—original draft, Writing—review & editing), Maria E. Yurgel (Conceptualization, Methodology), Samer Hattar (Supervision, Funding acquisition, Writing—original draft, Project administration, Writing—review & editing), and Jeff R. Jones (Conceptualization, Resources, Data curation, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review & editing)

Data Availability

Code and example data used in this review are available at https://doi.org/10.5281/zenodo.17632535