AI-Powered Mice Behavior Tracking and Its Application for Neuronal Manifold Analysis Based on Hippocampal Ensemble Activity in an Alzheimer's Disease Mice Model

Abstract

Investigating brain area functions requires advanced technologies, but meaningful insights depend on correlating neural signals with behavior. Traditional mice behavior annotation methods, including manual and semi-automated approaches, are limited by subjectivity and time constraints. To overcome these limitations, our study employs the YOLO neural network for precise mice tracking and composite RGB frames for behavioral scoring. Our model, trained on over 10,000 frames, accurately classifies sitting, running, and grooming behaviors. Additionally, we provide statistical metrics and data visualization tools. We further combined AI-powered behavior labeling to examine hippocampal neuronal activity using fluorescence microscopy. To analyze neuronal circuit dynamics, we utilized a manifold analysis approach, revealing distinct functional patterns corresponding to transgenic 5xFAD Alzheimer's model mice. This open-source software enhances the accuracy and efficiency of behavioral and neural data interpretation, advancing neuroscience research.

Keywords: Alzheimer’s disease; YOLO; artificial intelligence; behavioral classification; calcium imaging; locomotion; manifold; mice tracking; neuronal activity; neuronal network.

Figure 1

Schematic illustration of the experimental pipeline for AI-driven mouse position estimation and behavior scoring, integrated with neuronal manifold construction based on hippocampal activity recorded using a miniscope.

Figure 2

Mice tracking and core point estimation using the pretrained YOLO-Pose-v8 neuronal network. (A) Tracking mice position in the rounded arena using pretrained YOLO-Pose-v8. Red line is a boarder of the center zone, green line is a boarder of middle zone, blue line is a boarder of outer zone and pink line is a boarder of edge zone. (B) Loss curve of YOLO-Pose-v8 during training epochs. (C) Loss curve of YOLO-Pose-v8 during validation epochs. (D) Mouse in the rounded arena with highlighted core points by the MARS system. (E) Histogram distribution of error distances from manual mice core-point annotations. A red dotted line threshold of 0.25*D (PCK@0.25) is presented.

Figure 3

Determination of mice behavior type with pretrained YOLO-Pose-v11. (A) Composite frames approach for mice behavior scoring. (B) Loss curve of YOLO-Pose-v11 during training epochs. (C) Loss curve of YOLO-Pose-v11 during validation epochs for parameters fitting. (D) Accuracy of pretrained YOLO-Pose-v11 at identification mice behavior type. (E) Confusion matrix of pretrained YOLO-Pose-v11 at estimation of behavior. (F) Illustrative composite frames composition for distinct types of behavior as running, sitting and grooming, respectively.

Figure 4

Visualization of the mice movement and behavioral scoring results. (A) Mouse speed visualization for each frame—original (green line) and smoothed (red line). (B) Mouse velocity heatmap. (C) Movement distribution when freely moving in the rounded arena. (D) Position heatmap of mouse. (E) Mouse moving trajectory. All graphs illustrated single mouse behavior in the rounded arena.

Figure 5

Neuronal manifolds constrained on the miniature fluorescence microscopy data. (A) Schematic illustration of the neuronal manifold construction of the single recording based on the neuronal calcium traces. (B) Two-dimensional representation of the neuronal manifold architecture for WT+veh, 5xFAD+veh and 5xFAD+treat mice groups. (C) Significant decrease in the neuronal manifold intracluster distance is observed in the 5xFAD mice control group. WT+veh: n = 18 sessions, N = 7 mice; 5xFAD+veh: n = 18 sessions, N = 6 mice; and 5xFAD+treat: n = 17 sessions, N = 6 mice. Brown–Forsythe and Welch–ANOVA following Games–Howell’s multiple comparisons test was used (ns: non-significant, *: p < 0.05, ***: p < 0.001). All data are presented as the mean ± SEM.

Figure 6

Neuronal manifold geometrical features belonging to distinct behavior types are highly aberrant in transgenic 5xFAD mice. (A) Ellipses fitting for description of neuronal manifold architecture for running, sitting and grooming states. (B–D) Ellipse area constrained on the neuronal manifolds for running, sitting and grooming epochs, respectively. (E–G) Ellipses eccentricity based on the neuronal manifolds for running, sitting and grooming types of behavior, respectively. WT+veh: n = 18 sessions, N = 7 mice; 5xFAD+veh: n = 18 sessions, N = 6 mice and 5xFAD+treat: n = 17 sessions, N = 6 mice. Ordinary one-way ANOVA following Holm–Sidak’s post hoc test was used (*: p < 0.05, **: p < 0.01, ***: p < 0.001). All data are presented as the mean ± SEM.