Analysis of behaviour in the Active Allothetic Place Avoidance task based on cluster analysis of the rat movement motifs

The Active Allothetic Place Avoidance test (AAPA) is a useful tool to study spatial memory in a dynamic world. In this task a rat, freely moving on a rotating circular arena, has to avoid a sector where shocks are presented. The standard analysis of memory performance in the AAPA task relies on evaluating individual performance measures. Here we present a new method of analysis for the AAPA test that focuses on the movement paths of the animals and utilizes a clustering algorithm to automatically extract the stereotypical types of behaviour as reflected in the recorded paths. We apply the method to experiments that study the effect of silver nanoparticles (AgNPs) on the reference memory and identify six major classes of movement motifs not previously described in AAPA tests. The method allows us to analyse the data with no prior expectations about the motion to be seen in the experiments.

In what follows our proposed analysis method is first introduced and then applied to a data 114 set in order to demonstrate that it can successfully identify stereotypical types of behaviour 115 in the data. The observed differences in behaviour between treated and untreated animals 116 are then compared with standard analysis results to make sure that the results are consistent. Comparison of performance between untreated control (white) and treated (black) animals over a set of 5 sessions. Boxes represent the first, second (median, shown as a band) and third quartiles; whiskers are the minimum and maximum values. (a-d): Animals in the control group are able to quickly learn how to avoid the shock sector and perform on average much better than treated animals. Average speed (e) and the total length of the trajectories (f) do not show significant differences in locomotor activity between both groups.
Although the performance measures shown in Figure 1 identify a clear difference in per-154 formance in the spatial memory task between treated and non-treated animals, they provide 155 no indication about the types of behaviour that lead to such differences in the first place. The  Classification of trajectory motifs 160 The recorded trajectories (120 in total) for each animal and session were segmented (Mate-161 rials and Methods) resulting in 6,237 trajectory segments. A set of 11 features (Table 1)   because having a smaller number of dimensions is always desirable to avoid over-fitting, 188 = 6 was adopted here. For the target number of clusters = 6 was adopted since there is no 189 appreciable increase in the maximum correlation between clusters between = 5 and = 190 6, which suggests that = 5 contains one cluster that can be separated into two and that 191 = 6 is the more natural choice.  scription of the observed behavioural traits of each cluster is given below. Table 2 gives also 197 some statistics for each cluster, such as the relative size and percentage of segments that start 198 or end up within the shock sector.

Classes of behaviour 200
Here we describe briefly the behaviour associated with each of the resulting clusters. These  (Table 2). Relatively chaotic paths concentrated on the right/upper half of the arena, on the right side 214 and immediate vicinity of the shock sector.
Class 5: passive until shock sector, longer paths 216 Similar to class 2 but producing longer paths. Animals sit mostly in one position, frequently 217 completing a full revolution that starts and ends at the shock area (Table 2).    Figure 5. The Active Place Avoidance setup. Animals are placed on top of an elevated arena which is slowly rotating (1 revolution per minute). They can move freely around the arena but need to learn to avoid the shocks, which are delivered on sector, which is fixed according to the distal room cues. If they enter sector to be avoided, a short lasting low current pulse is delivered to their paws and repeated with a delay until they leave this sector. The position of the animals is tracked with LEDs 1 and 2, and a top-mounted camera.
Five recording sessions of 20 min each with a fixed shock sector (in the room coordinates) 319 were performed over a set of five consecutive days. This was followed by a test trial five days 320 later where the shock sector was not active. For the trials with an active shock sector animals 321 received a short (0.5 s) constant current pulse whenever they entered a predefined 60 ∘ shock 322 sector, which remained fixed across the trials. The amplitude of the shock pulses varied be-  In the present work, data analysis was also based on a classification of trajectory segments 347 into different stereotypical types of behaviour, however, classes of behaviour were not prede-348 fined. Also, no labelling of data of any kind was performed, that is, the classification performed 349 here was completely unsupervised.

350
The analysis done here consisted of the following steps:  Test animals have to move in the angular direction in order to evade the shock sector. There-370 fore, changes in the sign and magnitude of the angular speed were taken as the delimiting 371 points of the segments. More formally, in the second segmentation step trajectory points 372 were processed sequentially and added to a sub-segment until the difference between the 373 local and median angular speed of the sub-segment (recomputed for each new added point) 374 exceeded 0.6 rad/s. Segments shorter than 5 seconds were discarded in further analysis.

375
The two segmentation steps are shown schematically in Figure 6. From the original 120 376 trajectories, 1,741 segments were generated after the first segmentation step and 6,237 after 377 the second. Other statistics of the two segmentation steps can be seen in Table 3.    The 11 features computed for each trajectory segment were not used directly for clustering 418 the data. This is because in a high dimensional vector space the distance between elements 419 tends to be very similar, making it difficult to find meaningful clusters (Aggarwal et al., 2001).

420
In order to overcome this problem without having to explicitly select a small subset of features, Since no labelled data was used here our method is in practice completely unsupervised. Multi-factor testing of variance was done using a Friedman test (Siegel, 1956), a nonparamet-466 ric test that is well suited for data that is not normally distributed. The Friedman test is also a 467 matched test, and can control for experimental variability among subjects. In our case the 468 same animals were analysed over multiple sessions, which show a gradual change in be-