Standardizing the analysis of conditioned fear in rodents: a multidimensional software approach


  • P. Meuth,

    1. Institute of Physiology I, Westfälische Wilhelms-Universität Münster, Münster, Germany
    2. Department of Neurology – Inflammatory Disorders of the Nervous System and Neurooncology, Westfälische Wilhelms-Universität Münster, Münster, Germany
    Search for more papers by this author
    • These authors contributed equally to this work.
  • S. Gaburro,

    Corresponding author
    • Institute of Physiology I, Westfälische Wilhelms-Universität Münster, Münster, Germany
    Search for more papers by this author
    • These authors contributed equally to this work.
  • J. Lesting,

    1. Institute of Physiology I, Westfälische Wilhelms-Universität Münster, Münster, Germany
    Search for more papers by this author
  • A. Legler,

    1. Institute of Physiological Chemistry, University Medical Center of the Johannes Gutenberg Universität Mainz, Mainz, Germany
    Search for more papers by this author
  • M. Herty,

    1. Lehr- und Forschungsgebiet Mathematik – Kontinuierliche Optimierung, Rheinisch-Westfälische Technische Hochschule, Aachen, Germany
    Search for more papers by this author
  • T. Budde,

    1. Institute of Physiology I, Westfälische Wilhelms-Universität Münster, Münster, Germany
    Search for more papers by this author
  • S.G. Meuth,

    1. Institute of Physiology – Neuropathophysiology, Westfälische Wilhelms-Universität Münster, Münster, Germany
    2. Department of Neurology – Inflammatory Disorders of the Nervous System and Neurooncology, Westfälische Wilhelms-Universität Münster, Münster, Germany
    Search for more papers by this author
  • T. Seidenbecher,

    1. Institute of Physiology I, Westfälische Wilhelms-Universität Münster, Münster, Germany
    Search for more papers by this author
  • B. Lutz,

    1. Institute of Physiological Chemistry, University Medical Center of the Johannes Gutenberg Universität Mainz, Mainz, Germany
    Search for more papers by this author
  • H.-C. Pape

    1. Institute of Physiology I, Westfälische Wilhelms-Universität Münster, Münster, Germany
    Search for more papers by this author

Corresponding author: Dr S. Gaburro, Institute of Physiology I, Robert-Koch-Strasse 27a, D-48149 Münster, Germany. E-mail:


Data comparability between different laboratories strongly depends on the individually applied analysis method. This factor is often a critical source of variation in rodent phenotyping and has never been systematically investigated in Pavlovian fear conditioning paradigms. In rodents, fear is typically quantified in terms of freezing duration via manual observation or automated systems. While manual analysis includes biases such as tiredness or inter-personal scoring variability, computer-assisted systems are unable to distinguish between freezing and immobility. Consequently, the novel software called MOVE follows a semi-automatized approach that prefilters video sequences of interest for the final human judgment. Furthermore, MOVE allows integrating additional data sources (e.g. force-sensitive platform, EEG) to reach the most accurate and precise results. MOVE directly supports multi-angle video recordings with webcams or standard laboratory equipment. The integrated manual key logger and internal video player complement this all-in-one software solution. Calculating the interlaboratory variability of manual freezing evaluation revealed significantly different freezing scores in two out of six laboratories. This difference was minimized when all experiments were analyzed with MOVE. Applied to a genetically modified mouse model, MOVE revealed higher fear responses of CB1 deficient mice compared to their wild-type littermates after foreground context fear conditioning. Multi-angle video analysis compared to the single-camera approach reached up to 15% higher accuracy and two fold higher precision. Multidimensional analysis provided by integration of additional data sources further improved the overall result. We conclude that the widespread usage of MOVE could substantially improve the comparability of results from different laboratories.

New approaches in translational neuroscience (rodent to human relation) employ the behavioral analysis of laboratory animals as a tool for, e.g. (1) relating the contribution of a specific gene to a specific behavior or (2) testing novel drugs in the treatment of central nervous system-based pathologies. To this end, novel high throughput screening methods (rapid and reliable) are constantly developed to reduce data variability between laboratories.

However, the degree of such reproducibility has been shown to be very limited due to, e.g. varying breeding and housing conditions or experimenter's experience (Crabbe et al. 1999). Apparatus differences represent another source of interlaboratory variability in behavioral screening (Mandillo et al. 2008). Two independent laboratories investigating the same type of exploratory behavior in the open field revealed inconsistent results in three different mouse strains because of differing analysis methods. These differences were minimized after using a common method of analysis (Lipp et al. 2005).

Pavlovian fear conditioning (FC) has been extensively used for testing fear learning and memory performances as well as the efficacy of novel anxiolytic drugs. Thereby, an initially neutral conditioned stimulus (CS) is paired with an aversive unconditioned stimulus (US, e.g. a mild footshock). The CS can either be a specific cue, such as light or a tone (cued FC) or the environment (contextual FC; Phillips & LeDoux 1992). Explicit pairing(s) of CS–US elicit(s) the formation of a conditioned response including defensive behavior, autonomic arousal and stress (HPA) axis activation, which taken together to characterize a fearful state (Steimer 2002).

The fear level of rodents is typically quantified by assessing the duration of freezing periods which correspond to the absence of nonbreathing associated movements (Blanchard & Blanchard 1969). Depending on the applied protocol length, manual quantification can become very labor-intensive and thus gave rise to the development of several automated systems. Manual and automated analysis methods have been extensively studied and discussed previously (for review see Anagnostaras et al. 2010). In essence, the manual analysis is considered as the gold standard. However, it is subjected to unwanted human biases such as tiredness or a tendency to higher freezing scores if knowing the protocol or fur color of the animal. On the other hand, the fully automated freezing analysis of all currently available technologies (laser beam systems, video analysis and force-sensitive platforms) is not applicable for those animals with heightened unconditioned freezing (e.g. AJ mice; Hefner et al. 2008) as freezing could be easily confounded with immobility or sleeping stages.

Consequently, we thought of a semi-automatized system that combines the manual and automated analysis in a nonredundant and complementary manner but lacks their specific vulnerabilities. In order to be widely used and thus helping to reduce the methodological sources of variation, the system additionally had to provide essentially new functionalities.

The new platform is called MOVE which is the acronym for Multi data source Organization and Video Evaluation. It was tested for (1) validity in comparison to other established systems of freezing analysis and (2) its potential advantage for behavioral science.

Materials and methods

Description of the new platform: MOVE software system

The modular designed software system called MOVE comprises four basic elements that cover synchronous multi-angle video recording, manual key logging, automated video motion detection and multidimensional data integration (combination of multiple data sources).


The recording module operates (Fig. 1) with up to two video streams from standard laboratory video equipment, cost-effective webcams or a hybrid configuration. By using a standard office PC (Dual Core CPU 3 GHz, 4 GB RAM, Microsoft Windows 7, 64 bit) the software is capable of recording RGB (colored) videos from webcams with resolutions of up to 1600 × 1200 dpi for one stream or 800 × 600 dpi for two streams, respectively. The resulting .avi files are compressed via the freely available ‘Motion JPEG’ codec and thus compatible with standard external video players as well as the internal video player of MOVE.

Figure 1.

Combined view of the recording and key logger module. The left and right panel shows the views of the two webcams used to synchronously record the same object (mouse). Preliminary experiments revealed that this setup provides the most accurate results in freezing detection. The key logger function is shown in the bottom-center of the figure.

Key logging

The key logger (Fig. 1) is a module that is used to manually score different types of behaviors (e.g. freezing, grooming or sniffing). The user simply presses keyboard buttons to monitor the behavioral occurrences in any environment setting either online (during the experiment) or offline (analyzing video recordings). Termination of the logging results in an event list (Fig. 2) of observational data which can be replayed and reviewed to ensure accurate classification of the scored behaviors.

Figure 2.

Evaluation and integration module. This module resembles the core of the freezing analysis. Here, all data from the key logger, automated motion detection or different external data sources are brought together for the final integration. The upper part of the figure presents the motion detection graph that is analyzed according to a user-defined protocol and parameter set. In the lower part of the figure, the resulting event list is shown. Each entry can be replayed and edited to ensure the correct behavioral assignment.

Automated motion detection

Video files recorded by MOVE can be directly fed into the automated video motion detection which works with pixel differences on a frame-by-frame basis. The method of analysis comprises three different processing steps for two consecutive frames (Fig. 3). At first, the preceding frame (n – 1) is subtracted from the current frame (n) in order to remove or blacken all unchanged pixels (Kopec et al. 2007). The second step sums the intensity changes of all three color channels (red, green, blue; RGB) of that difference image and finally applies a user-defined minimum intensity threshold in order to strengthen the contrast. The third step is intended to filter out the so-called ‘salt and pepper noise’. For example, a single white pixel surrounded by solely black pixels will be blackened, as the moving object (rodent) always results in a consecutive amount of white pixels (for reference see 2D median filtering, Mathworks Matlab R2012a, medfilt2 and Fig. 3, right panel). Lowering the filter accuracy increases the analysis speed because fewer pixels have to be judged. The speed of the automated analysis process can be further increased by reducing the video area that has to be analyzed. This can be achieved by cropping the video area to a user-defined region of interest (ROI). This option additionally allows the user to exclude areas of the video that could impair the signal-to-noise ratio during the motion analysis (e.g. swinging cables required for electrophysiological recordings, see Fig. S1). Motion is quantified in terms of the total amount of nonblack pixels of the grey scale image resulting from the three previously described processing steps. On the basis of a user-defined maximum motion threshold MOVE finally combines all subthreshold time intervals within a user-editable event list (comparable to the key logger module) that can be replayed and reviewed.

Figure 3.

Automated motion detection. The left panel shows the currently analyzed video scene while the right panel highlights all changed/moved parts thereof with regard to the previous video time point. MOVE allows the so-called ‘batch processing’ thus analyzing a series of videos (experiments) without user interaction.

Data integration

Independent of the single data sources, normally used to assess a given behavior, there are situations or behavioral states where this type of measurement is limited in terms of resolution or signal-to-noise ratio. Consequently, MOVE offers a multi data source integration approach: after comparing the event lists of all different sources (e.g. video recordings, data from a force-sensitive platform, laser beam breaks) MOVE integrates all matching behavioral events into one list (logic AND). Event lists are standardized in order to support a large variety of different hardware devices and can be loaded from .xls files generated by Calc (OpenOffice) or Excel (Microsoft Office).

One example case could be the integration of two video recordings of the same animal from two different angles, as directly supported by MOVE. Depending on the animal's exact location and orientation a single camera could miss, e.g. slight horizontal head movements or sniffing (Fig. 4b, lower panel). This individual blind area (Fig. 4a: Cam 1 – blue zone, Cam 2 – green zone) can be fully covered by a second, oppositely located camera. Consequently the integration of both camera signals will reveal the animal's full movement spectrum.

Figure 4.

Camera positioning. (a) Two oppositely located cameras capture the animal's full movement spectrum by covering the blind area (Cam 1 – blue zone, Cam 2 – green zone) of the other camera. (b) Part b gives an example on mismatching camera results (Cam 1 – reports ‘movement’, Cam 2 – reports ‘freezing’) due to the blind area of camera 2.

Processing both camera signals with the automated motion detection module described before delivers a motion estimate (total amount of changed pixels) per video time point (Fig. 5a). The user-defined noise threshold (Fig. 5a: solid black horizontal line) separates the animal's real movements from noise sources like video flickering or small background changes. All video time points below that threshold are considered as freezing and consequently encoded as logic 1 in the corresponding binary time vector representation (Fig. 5b). Suprathreshold video time points are encoded as logic 0 (no freezing).

Figure 5.

Data integration example. (a) Part a depicts a two second cut-out (seconds 343–345) of the motion estimates derived by processing both camera recordings (camera 1 – blue trace, camera 2 – green trace) of Fig. 4 with MOVE's automated motion detection module. Motion is quantified in terms of the total number of changed pixels on a frame by frame basis (video time resolution: 15 frames/second). Manual freezing evaluation by an experienced observer (counterchecked by two additional blinded experts from the same laboratory) was used as reference (Ref, thick red lines indicate freezing intervals). The user-defined noise threshold (black horizontal line) separates the animal's real movements from potential noise sources. (b) On the basis of this noise threshold both video sources can be converted to their binary vector representations (α, β; 1 digit per frame, logic 0 – movement/above threshold, logic 1 – freezing/below threshold) and subsequently integrated via logic operators (γ – logic AND, δ – logic OR). The reference vector is set to logic 1 in case the manual observer reported freezing (thick, red line) and logic 0 otherwise. The accordance of a comparison vector with the reference vector can be quantified in terms of the total number of TP, TN, FN and FP digits (see also Table 1). (c) Given these values allows calculating common performance estimates [Prec – precision, Acc – accuracy, (Sen + Spec)/2 – (sensitivity + specificity)/2 or balanced accuracy]. Integration of two oppositely located cameras via logic AND (Ref vs. γ) clearly outperforms the single-camera approach (Ref vs. α, Ref vs. β) as well as the integration via logic OR (Ref vs. δ, left panel) independent of the applied performance estimate. Counting the number of matching (true: TP + TN) and mismatching (false: FP + FN) digits reveals the logic AND integration (Ref vs. γ) to reach a much higher overlap with the reference vector than the logic OR integration (Ref vs. δ, right panel).

The same recordings were manually analyzed by an experienced observer (Fig. 5a,b: Ref) to judge this algorithm's performance. Thereby, the presence of a thick red line marks a freezing time interval (logic 1) of the reference observer. A corresponding binary time vector representation can be derived as before.

Integration of several data sources or binary vector representations, respectively, is supposed to identify and logically solve mismatching statements. Figure 4b illustrates that in case of this freezing analysis example mismatches are solely caused by the blind area of each single camera. Here, camera 2 states freezing (logic 1, lower panel) while camera 1 states movement (no freezing, logic 0, upper panel). This mismatch must be solved by the logic AND operator (Table 1) in order to guarantee that one movement statement (logic 0) overrules all others. Using the logic OR operator (Table 1) for example would invert this logic in a way that a single freezing statement (logic 1) overrules a second or even more cameras stating movement (logic 0) and is thus inapplicable.

Table 1. Terminology definition
 Reference vectorComparison vectorANDOR
  1. Comparing two binary time vector representations (reference and comparison vector) digit per digit allows analyzing their overall concordance. Thereby, each single digit comparison results in an either TN, FP, FN or TP event. For instance, if the reference vector states freezing (logic 1) while the comparison vector does not (movement, logic 0) the currently regarded time point will be assessed false negative (FN). Boolean algebra comprises a multitude of different functions for the integration of binary vectors with logic AND and logic OR being the most basic ones among them.

Performance estimates

Given the total amount of matching [true positive (TP), true negative (TN)] and mismatching [false positive (FP), false negative (FN)] digits after comparing two binary vectors allows calculating common performance estimates:

display math

This value indicates the closeness of the measurement to the reference value (e.g. human observer).

display math

This value indicates the replicability of the measurement (Joint Committee for Guides in Metrology 2008).

display math

This value indicates the likelihood of positive events to be found.

display math

This value indicates the likelihood of negative events to be found.

display math

This value equilibrates misleading sensitivity and specificity trends on imbalanced data sets.

Referring to Fig. 5, these performance estimates were exemplarily used to contrast the integration of both camera signals via the logic AND operator (γ) with that via the logic OR operator (δ) as well as the single source approaches (α, β). Sensitivity and specificity are known to produce inflated performance estimates on imbalanced data sets and were thus combined (balanced accuracy). A corresponding example is shown in Table 2. It can be concluded that the data integration of both camera signals via the logic AND operator (Ref vs. γ) clearly outperforms a single-camera approach (Ref vs. α, Ref vs. β) independent of the applied performance estimate (Fig. 5c, left panel). This does not hold for the data integration via the logic OR operator (Ref vs. δ). In terms of the total number of matching (true: TP + TN) and mismatching (false: FP + FN) digits the logic AND integration (Ref vs. γ) shows a larger compliance to the reference vector (Fig. 5c, right panel) than the logic OR integration (Ref vs. δ).

Table 2. Inflated performance estimates on imbalanced data sets
  1. Comparing the binary vectors A and B with the reference vector (Ref) results in misleading sensitivity (sens) values in the upper and specificity (spec) values in the lower example with regard to the total number of matching digits (matches).

Inter- and intralaboratory variability in freezing detection

Freezing is an innate defensive behavior defined as complete immobility with the exception of respiratory movements. Differing analysis techniques as well as individual human influences (e.g. experience, tiredness, expectancy) gave rise to the question whether or not the comparability of freezing results between different laboratories and even within the same laboratory might consequently be impaired.

Therefore, video material provided by six external laboratories (Fig. 6a–f) was manually reanalyzed by a well-trained local observer blind to the individual experiment background (conditioning protocols, animal strain, working hypothesis, evaluation technique) as well as the external evaluation results. The same conditions were applied for a second, independent observer of the same laboratory reanalyzing in-house video recordings in order to assess the intralaboratory freezing detection variability. All videos were free of visual or acoustic information able to directly or indirectly influence the manual scorer (bias of expectancy).

Figure 6.

Intra- and interlaboratory variability. Freezing detection differences between a local well-trained observer (manual analysis) and another well-trained observer (manual analysis) of the same laboratory (videos analyzed: n = 6, labeled ‘Int’ for internal) or other analysis systems employed in 6 other research groups (videos analyzed: n = 2, 3, 5, 9, 8, 12, labeled with letters A, B, C, D, E, F respectively) based on single source video material. One-way anova revealed differences among interlaboratory analysis F(5, 111) = 17.983, P < 0.001. Tukey post hoc results are indicated as follows: **P < 0.01 Lab A vs. Lab E and Lab F, ###P < 0.001 Lab B vs. Lab E and Lab F, §§§P < 0.001 Lab D vs. Lab E and Lab F. The shaded background area represents the range of freezing analysis variation that is generally considered acceptable.

Analyzed experiments comprised different animal strains (C57Bl/6J mice, Sprague Dawley rats), conditioning protocols (cued and context FC), evaluation phases (extinction and recall of extinction), scoring methods (manual, video software and laser beam analysis) in order to reach maximum heterogeneity or explanatory power, respectively (Table 3). A minimum freezing event duration of 1 second was constantly applied.

Table 3. Data overview
  1. Calculating the intra- and interlaboratory variability in freezing detection comprised C57Bl/6J mice and Sprague Dawley rats. Animals were initially fear conditioned (cued and context FC) and later on scored at two distinct phases (extinction and recall of extinction). External scoring methods included manual, video software (VS, Med Associates Inc.) and laser beam (LB, TSE Systems GmbH) evaluation. Total number of animals (number) varied due to laboratories' availability. Compared were the mean freezing score (%, freezing time vs. protocol length) of the local reference observer (reference) with that of the external laboratories (comparison).
  2. Int, intralaboratory comparison; A–F, interlaboratory comparison.
ASprague DawleyCuedExtinctionManual232.4542.33
DC57Bl/6JCued + contextExtinctionVS934.9240.20
EC57Bl/6JCuedExtinction + recallLB89.2638.54

The averaged freezing scores reflect the overall freezing duration in relation to the protocol length and are thus given as percentage (%). Inter- and intralaboratory freezing detection variability (delta freezing, Fig. 6) was calculated as follows (L: local reference observer, A: another observer of the same local laboratory, E: observer/validated analysis system of an external laboratory):

display math

Positive or negative delta values correspond to higher or lower freezing scores of the local reference observer, respectively.

Fear conditioning

Multidimensional data analysis

In contrast to the standard, single data source recordings provided by the cooperating external laboratories (Figs 6 and 7) a separate cohort of 10 animals was fear conditioned and evaluated (Figs 9 and 10) using the full data spectrum of the local, multidimensional recording test setup (3D: force-sensitive platform combined with two webcams).

Figure 7.

Reliability of MOVE. Compared were MOVE evaluation results (a–f, black bars) with the results (a–f, white bars) derived by the externally used evaluation techniques based on single-source video material (Table 3). STT: unpaired Student's t-test, ANO: one-way anova, TPH: Tukey post hoc, HF: high freezing and LF: low freezing animals. Comparison results: (a) STT: t38 = 0.949, P = 0.3487. (b) STT: t4 = 0.287, P = 0.788. (c) STT: t18 = 1.213, P = 0.253. (d) ANO: F(7, 111) = 62.42, P < 0.001; TPH: P = 0.783. (e) ANO: F(3,79) = 15.15, P < 0.001; TPH: P = 0.685. (f) STT: t16 = 0.432, P = 0.672. In summary no significant differences could be detected.

Experiments were carried out with adult male C57Bl/6J mice (8 weeks, Charles River, Germany). Animals were singly housed at least 7 days prior to commencement of any experiment, kept in a 12:12 h light–dark cycle (lights on at 7 am) and provided food and water ad libitum. All procedures were performed under strict observance of the European Committees Council Directive (86/609/EEC) for experimentation in animals and were approved by the Bezirksregierung Münster (LANUV NRW, AZ 87-51.042010.A189). FC was performed according to a previously established protocol (Lesting et al. 2011). Briefly, on day 0 mice were habituated for 6 min to six tone presentations (unconditioned stimulus, CS, 2.5 kHz, 85 dB, 20 seconds) with a regular intertrial interval (ITI) of 30 seconds. Two sessions were performed within 1 day but 6 h apart. On day 1, fear training took place consisting of three tone presentations (conditioned stimulus, CS+, 10 kHz, 85 dB, 10 seconds, randomized ITI between 10 and 30 seconds) each co-terminated with a mild footshock (0.4 mA, 1 second). On the next day, mice were put in a neutral context (clear plastic box, L × W × H: 19 × 19 × 13 cm, filled with sawdust bedding) to conduct the fear retrieval session (6 min duration) during which both the CS and CS+ were presented without footshock (4 × 10 seconds CS at 20 seconds ITI, 40 seconds no tone, 4 × 10 seconds CS+ at 20 seconds ITI).

CB1 deficient mice

To check the applicability of MOVE to genetically modified mouse models a total of eight cannabinoid receptor type 1 deficient mice (CB1–/–; Marsicano et al. 2002) and eight wild-type littermates (CB1+/+) were exposed to foreground context FC.

Animals were singly housed 1 week prior to the onset of the experiment and exposed to conditioning on day 7. The conditioning protocol comprised 198 seconds in the conditioning chamber (metal grid, lights on) before the application of a mild footshock (0.7 mA, 2 seconds). Animals were allowed to explore the conditioning chamber for additional 60 seconds and were then returned to their home cage. On the next day (24 h later), a re-exposition in the same context and a 5 min lasting fear expression evaluation followed.

Statistical analysis

All group data are presented as mean ± SEM (standard error of the mean) and were tested for homoscedasticity using Levene's test (Statistica software). Statistical significance was calculated via one-way anova and Tukey post hoc in case of multiple group comparisons and unpaired Student's t-test otherwise. Results are considered significant at P < 0.05 and further distinguished as follows: *P < 0.05, **P < 0.01, ***P < 0.001.

Results and discussion

MOVE strongly differs from other currently available products in three major points:

  1. Manual (Fig. 1) and automated video analysis (Fig. 3) are combined within one software.
  2. Synchronized recordings can be obtained from two independent video sources (Fig. 1) supporting cost effective webcams, standard laboratory equipment and even the mixture of both.
  3. Multidimensionality via the integration of arbitrary data sources (Figs 2 and 5) allows reaching the highest possible accuracy and precision (Figs 9 and 10) in data analysis.

This study provides evidence for intra- and interlaboratory variations in freezing evaluation, the reliability and applicability of the MOVE software system in comparison to established alternative analysis methods as well as the significant accuracy and precision improvements gained by multidimensional data analysis.

Intra- and interlaboratory variability

Calculating the intralaboratory variability was conducted by comparing the freezing analysis of a local reference observer to the analysis of another well-trained and experienced member of the same laboratory. Interlaboratory variability was derived by comparing the analysis of the same reference observer to that of six external laboratories characterized by a long lasting FC expertise and a comprehensive methodological spectrum. Human scoring variations of about 10% are a commonly accepted threshold (Fig. 6, shaded area), which accounts for biases such as tiredness, anticipation or slight perception differences (Anagnostaras et al. 2010; Brunzell et al. 2002). In this regard, the intralaboratory variability (Int, 6.28 ± 2.03%) as well as the comparison to four out of six external laboratories (Lab A –9.88 ± 5.06%, Lab B 4.57 ± 6.07%, Lab C –10.44 ± 7.58% and Lab D –5.28 ± 1.66%) were within the above mentioned range. Different results were obtained for Lab E (–29.28 ± 1.48%) and Lab F (–36.40 ± 4.83%) showing up to 36% higher freezing scores (Fig. 6, P < 0.001) as compared to the local reference observer. These data clearly demonstrate that the manual analysis of freezing behavior can significantly differ among laboratories and thus limits the overall result comparability. This is in good agreement with similar findings previously reported for other types of behavior (Bonasera et al. 2008).

Reliability of the MOVE software system

MOVE follows a semi-automatized approach that splits the overall video analysis in two processing stages. The first stage automatically identifies the video sequences of interest in terms of quantified motion, whereas the second stage resembles a quality assurance during which the manual observer finally judges all preselected video sequences. By disburdening the manual observer from the most monotonous and time-consuming parts of the video analysis, the influence of biases such as tiredness or poor concentration can be clearly reduced.

In order to the check this approach all videos provided by the six cooperating laboratories were re-analyzed using the MOVE software system. Comparing the MOVE supported freezing results to those of the external laboratories (Fig. 7a–f) revealed no significant (P > 0.05) differences any longer. Consequently the use of MOVE allows compensating manual analysis differences of up to 36% (Fig. 6). In light of the various analysis methods (Díaz-Mataix et al. 2011; Duvarci et al. 2011; Hefner et al. 2008; Jasnow et al. 2009; Morawska & Fendt, 2012; Reger et al. 2012; Seidenbecher et al. 2003) used to derive the external freezing evaluation, MOVE additionally proved to be as reliable (Fig. 7, no significant differences) as all other currently available analysis systems (Table 3) while dealing with single data sources.

Applicability to genetically modified mouse models

In order to prove its general applicability and impact on behavioral sciences, we used MOVE to evaluate video material from the well-established cannabinoid receptor type 1 (CB1) deficient mouse model (CB1–/–; Marsicano et al. 2002). For cued FC, it was shown that CB1–/– animals and their wild-type littermates (CB1+/+) did not differ in their initial freezing response, but with continuous CS presentation it was apparent that CB1–/– animals were impaired in within-session extinction, whereas CB1+/+ animals did normally extinct (Marsicano et al. 2002; Plendl & Wotjak 2010). Regarding the contribution of CB1 in contextual fear memory formation, recent studies showed contradictory results. In a foreground contextual FC task, CB1–/– mice on the CD1 background responded with lower fear levels to the context as compared to CB1+/+ animals (Mikics et al. 2006). In contrast, Jacob et al. (2012), performing background contextual FC with CB1–/– mice on the C57Bl/6N background, found enhanced freezing levels in the conditioning context after a high-intensity conditioning foot shock of 1.5 mA.

In this study, eight CB1–/– animals and an equal number of CB1+/+ controls were exposed to foreground FC (see Materials and methods, CB1 deficient mice). Conducting all evaluation steps with the MOVE software system revealed significantly (P < 0.01) higher fear responses for the CB1–/– as compared to the CB1+/+ animals (Fig. 8). This is consistent with the data of Jacob et al. (2012), who used the same genetic background, a similar protocol, but higher shock intensity (1.5 vs. 0.7 mA).

Figure 8.

Applicability of MOVE. Evaluating the fear response (foreground context FC) of cannabinoid receptor type 1 deficient mice (CB1–/–, n = 8) and their wild-type littermates (CB1+/+, n = 8) with MOVE revealed significant (unpaired Student's t-test: t14 = 3.161, **P < 0.01) differences.

Figure 9.

Multidimensional data analysis: 2D. Integrating two cameras (a,b, black bars) significantly improves the freezing behavior analysis results of a single camera system (a,b, white bars) in terms of accuracy (a, unpaired Student's t-test: t18 = –2.892, **P < 0.01) and precision (b, unpaired Student's t-test: t18 = –2.823, *p < 0.05). Mice (n = 10, C57Bl/6J) underwent auditory FC. The manual reference was counterchecked by two additional blinded experts of the same laboratory.

Figure 10.

Multidimensional data analysis: 3D. Integrating data from a force-sensitive platform with data from two cameras (a,b, black bars) significantly improves the freezing behavior analysis results of the exclusive platform approach (a,b, white bars) in terms of accuracy (a, unpaired Student's t-test: t18 = –2.281, *P < 0.05) and precision (b, unpaired Student's t-test: t18 = –2.181, *P < 0.05). Mice (n = 10, C57Bl/6J) underwent auditory fear conditioning. The manual reference was counterchecked by two additional blinded experts of the same laboratory.

Advantages of multidimensionality

While sharing the reliability and applicability of other analysis systems in the one-dimensional data analysis MOVE's real advantage can be shown when it comes to multidimensionality. One of its core features is the option to generate highly synchronous video recordings of the same object from two different perspectives independent of the used video hardware and integrate them with other arbitrary data sources. Given this new technology we added two webcams to an already existing setup, which records animal movements via a force-sensitive platform, in order to probe whether three independent data sources used in parallel would increase the sensitivity of the freezing detection. Thereafter, this configuration was used to record and analyze the freezing behavior of 10 fear conditioned mice (Lesting et al. 2011) with the aid of MOVE.

Video recordings were carried out with both webcams oppositely located on the long side of the rectangular macrolon type II cage (Fig. 4). This approach greatly improves the vicinity to the recording object in comparison to a top view camera and covers the object's full movement spectrum (Fig. 4, no blind areas). Furthermore, this camera positioning reduces disturbances by, e.g. mounting cables of electrophysiological or optical setups (Fig. S1). Using the semi-automated video analysis of MOVE resulted in a single event list per video file (angle/perspective). The two recordings/event lists of the same object were integrated by applying logic AND to the corresponding binary time vectors (Fig. 5).

Comparing this integrated signal to the manual scoring of the human observer (reference) allowed calculating the total amount of TP, TN, FP and FN time points (see Materials and methods, data integration) and consequently the two commonly used performance estimates accuracy and precision (see Materials and methods, performance estimates). Thereby, the dual-camera approach significantly improved the single camera results by 10.78 ± 3.72% accuracy (Fig. 9a, P < 0.01) and 19.05 ± 6.75% precision (Fig. 9b, P < 0.05).

Data from the force-sensitive platform represents a movement estimation per time point comparable to the evaluated video recordings depicted in Fig. 5a. Application of a user-defined threshold distinguished time points of solely breathing-associated movements (logic 1) from phases of higher activity (logic 0) and thus lead to a corresponding binary vector representation. Integrating (logic AND) these vectors with the combined video vectors (2 camera results) revealed an average accuracy and precision improvement of 5.23 ± 2.29% (Fig. 10a, P < 0.05) and 18.06 ± 8.27% (Fig. 10b, P < 0.05), respectively in comparison to the platform results alone.

Limitations and alternatives

Automated freezing behavior analysis (including the automated motion detection module provided by MOVE) often fails in distinguishing freezing from other immobility states (e.g. sleep). This issue can be generally solved by invasive technologies such as the surgical implantation of telemetric electrocardiogram (ECG) devices for the real time detection of an increased heart rate during fear stages (Gaburro et al. 2011).

However, surgical interventions could be neither desired nor applicable in a given experimental situation such that the manual analysis resembled the only alternative so far. The new, semi-automated approach of MOVE resembles an advantageous compromise between both solutions. Using its multidimensional data integration provides the user the most accurate, rapid and unbiased pre-selection of potential freezing video sequences for the final human judgment. Even in case the user has to opt for a completely manual analysis (e.g. low signal-to-noise ratio or lack of additional data sources) the integrated key logger of MOVE will be of further help.


Multidimensional data integration provides a significant advantage in terms of accuracy and precision in freezing behavior analysis. Since this significance could already be reached by adding a single data source any existing setup should be easily upgradeable. We recommend using MOVE for three major reasons. First of all MOVE supports synchronized recordings from standard laboratory video equipment as well as cost-effective webcams making it fast and affordable to go for one or two additional video data sources at any behavioral neuroscience setup. Second, its standardized event list interface allows integrating all kinds of different data sources (no hardware constraints) thereby enabling a significantly more accurate and precise multidimensional data analysis. Considering also the integrated key logger and video player, MOVE offers a complete all-in-one solution that prevents any data incompatibility or conversion effort necessary while working with different products. On the basis of that, a third argument for MOVE is its great potential for a widespread usage that in turn contributes to higher inter- and intralaboratory result comparability.


The authors would like to thank Michael Fanselow, Aaron Jasnow, Patrick Cullen, Carsten Wotjak, Nicolas Singewald, Nigel Whittle, Denis Paré, Sevil Duvarci, Markus Fendt, Joseph Ledoux, Lorenzo Diaz-Mataix, Andrew Holmes, Ozge Gunduz-Cinar and Shaun Flynn for the provided video material, constructive suggestions and comments for the validation and improvement of MOVE. The project was supported by SFB-TRR58, TP A02 (to T.S. and H.-C.P.), TP A04 (to H.-C.P. and B.L.) and FOR926 (to B.L.). Requests from academic and non-profit organizations should be addressed to the corresponding author. Commercial entities should contact the copyright holder Clinic Invent ( The authors declare no conflict of interest.