Automatic recognition of explorative strategies in the Morris water maze

https://doi.org/10.1016/S0165-0270(03)00187-0Get rights and content

Abstract

Notwithstanding the development of reliable tracking systems, the quantification methodology of the Morris water maze (MWM) has witnessed an operational mismatch between the indexes used to quantify MWM performance and the cognitive concepts derived from these indexes. Indeed, escape latency is the main, and often unique, performance measure used for the quantification of behavior. Aim of the present work was to overcome this limitation by presenting a methodology that allows for automatic categorical pattern recognition of the behavioral strategies performed in the MWM. By selecting few a priori and user-defined behavioral categories, many quantitative variables and regions of interest (ROIs), we used discriminant analysis (DA) to obtain 97.9% of correct automatic recognition of categories. The developed discriminant model (DM) also allowed to predict category membership of newly recorded swim paths with the same statistical efficacy (96%), and to identify the variables that better discriminate between adjacent categories. The combination of DA with a tracking system, a selection of many variables, different ROIs and qualitative categorization, reduces the gap between the measurement process and the categories used to describe a given behavior, and offers a methodology to computationally reproduce the human categorization of behaviors in the MWM.

Introduction

The Morris water maze (MWM), described about 20 years ago as a behavioral device to investigate spatial learning and memory in laboratory rats and other species (Morris, 1981, Morris, 1984, Morris et al., 1982), has become one of the most frequently used laboratory tools in behavioral neuroscience. Briefly, the MWM consists of a large circular pool filled with opaque water in which a small escape platform is hidden. During a number of training trials, the animals learn to find the platform and to escape from the water. Although the basic procedure is relatively simple, the MWM has been used in some of the most sophisticated experiments on the neurobiology and neuropharmacology of spatial learning and memory to validate rodent models for neurocognitive disorders and to evaluate possible neurocognitive treatments (Morris et al., 1986, Brandeis et al., 1989, McNamara and Skelton, 1993, Morris and Davis, 1994, D'Hooge and De Deyn, 2001 for a review).

However, despite its imposing credentials, MWM quantification methodology has not witnessed parallel development. Thus, there is an operational mismatch between the indexes used to quantify MWM performance and the cognitive concepts derived from these indexes. Notwithstanding the development of reliable tracking systems to obtain precise measures of an animal's performance in terms of the spatial coordinates in time, this methodological gap seems to be neglected by the majority of researchers working with the MWM. Indeed, very basic and quick indexes are preferred to more complicated and time-consuming quantifications, resulting in an oversimplified description of the behavior.

Standard performance measures usually include escape latency and length of the swimming path. Lindner (1997) argued that path length might be the most appropriate index of cognitive performance in the MWM. Furthermore, several authors (Gallagher et al., 1993, Stewart and Morris, 1993, Dalm et al., 2000) have recommended the use of alternative measures, such as swim speed, path directionality, cumulative distance to platform, proximity, path tortuosity, and turning preferences. Other measures can be derived from a selective analysis of a target zone, to obtain indexes such as time spent in target zone, distance to target, number of crossings of a target, and so on.

All of these indexes face two basic problems when applied to the actual performance of the animal in the MWM. First, although each index is correlated with some aspects of performance and often considered representative of the performance, it does not describe the actual performance itself. Second, even the simultaneous utilization of different indexes does not provide any “integrated” information about a behavior, but simply sums different aspects of the behavior. Nevertheless, most of the literature on the MWM is based on escape latency as the main, and often unique, performance measure used for the quantification of behavior.

Another approach to study MWM performances relies on the use of manual behavioral categorizations (Petrosini et al., 1996, Wolfer and Lipp, 2000). This methodology allows a further step in the analysis of performance since it provides a qualitative description of complex explorative behaviors. In fact, very accurate behavioral categories can be obtained by observing the actual swimming behavior of the animal. However, being a qualitative, and not quantitative approach, this time consuming methodology can be sometimes unreliable.

An attempt to overcome these limitations was recently made by Lipp and Wolfer (Lipp and Wolfer, 1998, Wolfer and Lipp, 2000, Wolfer et al., 1998, Wolfer et al., 2001). They proposed applying factor analysis (FA) to many quantitative indexes. FA is a bottom–up (from data to concepts) inferential statistical method that makes a dimensionality reduction of the sample by extracting a few “artificial” variables, so called factors, that can be more or less correlated with each original variable. However, the extracted factors are still far from being a categorical representation of the behaviors observed. Indeed, on the basis of its high correlation with some of the variables, a factor can be labeled as “thigmotaxis”, “retention”, “passivity” or “chaining” and so on, but does not provide any categorical membership of the individual swim path.

The aim of the present work was to overcome this limitation by presenting a methodology that allows for automatic pattern recognition of the behavioral strategies performed in the MWM. To this purpose we used discriminant analysis (DA), a very common statistical method in psychology, that allows differentiating two or more a priori defined and multivariate groups (Cooley and Lohnes, 1971, Gnanadesikan, 1977, Dillon and Goldstein, 1984, Cinanni, 1990) (Fig. 1). DA permitted identifying the response profile that best discriminated between groups, predicting new data membership and determining which variables are the best predictors of group membership. By selecting few a priori and user-defined behavioral categories, many quantitative variables and regions of interest (ROIs), we performed automatic categorical pattern recognition in the MWM.

Section snippets

Animals

Swimming paths (tracks) were collected from a total of 37 adult male Wistar rats. Thirty of them were naı̈ve animals and seven were adult hemicerebellectomized rats (250–300 g). They were housed two animals per cage with free access to food and water and standardized dark/light schedule (10-h dark/14-h light).

The HCb was performed to compare tracks of normal animals with those of HCbed animals known from previous studies to exhibit peculiar explorative strategies (Petrosini et al., 1996,

Automatic recognition of the categories

When the a priori prediction was performed on the previously categorized data set of 1049 cases using all seven categories and 28 variables at a time, we obtained an automatic correct recognition of 97.9% of cases (Table 2). Out of 1049 swim paths, we obtained a total of 22 misclassified paths. Of these errors, two were misclassifications in a non-adjacent category (serious errors), whereas 20 were misclassifications in adjacent categories (slight errors). This model was subsequently used for

Discussion

The main finding of the present paper is to have rendered quantitative the categorical analysis of MWM behavior by means of an automatic “gestaltic” recognition of a priori and user-defined categories of cognitive strategies. Co-processing of many quantitative variables allows sophisticated analyses of MWM behavior, offering a methodology to computationally reproduce a human categorical process.

The combination of DA with a tracking system, a selection of many variables, different ROI and

References (25)

  • V. Cinanni

    Dimensioni di somiglianza

    (1990)
  • W.C. Cooley et al.

    Multivariate Data Analysis

    (1971)
  • Cited by (84)

    • Differential navigational strategies during spatial learning in a new modified version of the Oasis maze

      2020, Behavioural Brain Research
      Citation Excerpt :

      The problem of interpreting the meaning of those factors or latent variables is a well-known problem in factor analysis. Graziano et al. [20] addressed this problem by classifying the animal behavior in several qualitative categories. Then they performed a discriminant analysis over a comparable list of variables.

    View all citing articles on Scopus
    View full text