Machine learning-based model for behavioural analysis in rodents applied to the forced swim test

Figures

• 

  Fig. 1. Historical count of the 'forced swim test’ tag in PubMed. From 2000 to 2024, a PubMed search retrieved more than 8000 scientific articles using 'forced swim test’ as a keyword. In particular, 5500 articles have been published in the past 10 years alone.

• 

  Fig. 2. Example of 15 frames extracted from the 3D-tensor, for each behaviour. Each 3D-tensor contains 75 frames. For the present representation, one frame was shown every five frames.

• 

  Fig. 3. The model architecture used for behaviour classification. The model is composed of a cascade of 3D convolutions organized in residual blocks to extract the spatio-temporal features of the data. Then, after an average global pooling, a series of dense layers was used to learn non-linear combination of the extracted features. Finally, a three-node softmax layer is used to produce a probability distribution of the possible behaviours label. The behaviour corresponding to the highest probability is the output of the model. We set the kernel size of the 3D convolutions to (3,3,3), the kernel size of the 3D average pooling to (2,2,2), and the dropout value to 0.1. The variable n in the convolutional neural layer is the number of its output channel.

• 

  Fig. 4. Performance of the ML scoring compared with manual scoring. The confusion matrix on the validation set (A). The confusion matrix on the sum of the drug experiment videos (B). As an example, a video has been selected to show the concordance between the two scoring methods via the confusion matrix (C) and via the time plot in which any mismatch between the two methods is also shown (red bars). Each 5-min video is split into a sequence of 100 3-s tensors, used as inputs for the ML algorithm (D).

• 

  Fig. 5. Comparison between the behaviour times of behaviour measured with the two scoring systems. Scatter plots of the percentage time performing the three behaviours predicted by the trained model show the positive correlation between manual and ML labeling systems. As measured by a Pearson’s correlation, the following scores are obtained: immobility \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r=0.918$$\end{document}; swimming \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r=0.858$$\end{document}; climbing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r=0.955$$\end{document}.

• 

  Fig. 6. Measurement of the effect of the drugs on FST behaviour measured with ML algorithm and manual scoring. Comparison between the manual and the ML algorithm scoring following drug treatment. When the ML algorithm was compared with manual scoring, no statistically significant differences were detected between the two methods for any behaviour assessed in the FST (Immobility: F(1,42) = 0.31, p = 0.579; Swimming: F(1,42) = 1.5, p = 0.227; Climbing: F(1,42) = 1.64, p = 0.208). Consistent with well-established literature on the effects of FLX (n = 8) and DMI (n = 8) in the FST, both methods successfully identified: (A) a decrease in immobility time (F(2,42) = 19.42, p < 0.0001) in rats treated with antidepressants compared to the vehicle group (Manual: vehicle vs. DMI, p < 0.0001; vehicle vs. FLX, p = 0.0003. ML: vehicle vs. DMI, p = 0.007; vehicle vs. FLX, p = 0.003) (B) the SSRI FLX preferentially increased swimming behaviour (F(2,42) = 13.34, p < 0.0001. Manual: vehicle vs. FLX, p = 0.002. ML: vehicle vs. FLX, p = 0.018); and (C) the TCA DMI preferentially increased climbing behaviour (F(2,42) = 13.02, p < 0.0001. Manual: vehicle vs. DMI, p = 0.0003. ML: vehicle vs. DMI, p = 0.008).…

• 

  Fig. 7. Measurement of the effect of the drugs on FST behaviour measured with ML algorithm. The ML algorithm accurately categorizes the effect of different classes of antidepressants in the FST. According to the literature, our ML model assigned the TCA AMI (n = 5) high levels of climbing, whereas, for the SSRI PRX (n = 5), the model detected increased swimming. In the case of VLX (n = 5), which, compared to TCAs and SSRIs, has a more balanced noradrenergic/serotonergic profile, our model indicated a corresponding balance between swimming/climbing performance. (A) Immobility (F(3,16) = 4.48, p = 0.018. vehicle vs. AMI, p = 0.014; vehicle vs. PRX, p = 0.656; vehicle vs. VLX, p = 0.044); (B) Swimming (F(3,16) = 5.93, p = 0.006. vehicle vs. AMI, p = 0.97; vehicle vs. PRX, p = 0.027; vehicle vs. VLX, p = 0.011); (C) Climbing (F(3,16) = 10.55, p = 0.0005. vehicle vs. AMI, p = 0.023; vehicle vs. PRX, p = 0.075; vehicle vs. VLX, p = 0.541). One-way ANOVA followed by Dunnett’s multiple-comparisons test was used. Data are presented as mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs}…

• 

  Fig. 8. Dataset construction. To build the dataset the video is split into 3-s sub-videos and each sequence of frames composes a single data; every frame is subtracted from the first frame of the video containing no rodent; then, the ROI (blue line) of the cylinder is wrapped into a rectangle of standard dimension (LxH). Finally, the resulting 3D-tensor of dimension TxLxH is normalized. The related portion of the video was manually evaluated by two annotators to recognize immobility, swimming, and climbing behaviours (A). The final dataset is composed of the three interest behaviours (B).

• 

  Fig. 9. Pipeline of ML annotation applied to a video. The video is split into 3-s sub-videos, and they are preprocessed as already described in Fig. 8: every frame of the sequence is subtracted from the first frame of the original video containing no rat; then, the ROI (blue line) of the cylinder is wrapped into a rectangle of standard dimension (LxH). Finally, the resulting 3D-tensor of dimension TxLxH is normalized. The resulting 3D tensor is evaluated by the model already trained.