End-to-end pipeline for quantifying rat behaviour in open-field and T-maze recordings using DeepLabCut pose estimation, rule-based behaviour detection, spatial-occupancy analysis, and a downstream ML classifier stack (LOOCV + SHAP).
This repository accompanies a thesis on per-subject and per-cohort behavioural profiling (MA1 / MA3 / MA5 / MA7).
- Open-field arm: end-to-end pipeline complete. 12 subjects (4 cohorts × 3 rats)
processed through DLC → behaviour detection → spatial analysis → feature
engineering → LOOCV-validated ML classifiers (logistic / RF / SVM / XGBoost
across 4 targets) → SHAP interpretation. See
docs/final_report.md. - T-maze arm: reactivated 2026-05-05 after a brief deferral. Now in
planning: a separate, lighter 5-point DLC model (vs. the OFT 9-point
set) — small subjects, corridor wall occlusion, and top-down-only camera
made the OFT model unsuitable. Keypoint set, per-behaviour metric mapping,
and repo layout decisions live in
docs/tmaze_keypoints_and_layout.md. - Cohort statistics (replaces failed cohort classifier):
Kruskal-Wallis + Dunn post-hoc + PERMANOVA pipeline at
analysis/cohort_stats.py— main result:center_zone_entriesshows a large effect (ε²=0.69) with Grapefruit < ASP+GF (Dunn z=2.85, within-feature q=0.026); other anxiety-axis features point the same direction but don't survive BH-FDR. Outputs inreports/cohort_*.csv. - Next planned model — window-level behaviour classifier: the
subject-level ML stack hit n=12 limits (cohort F1≈0.27, label leakage on
anxiety_level). Plan now is to train at the window level (1-second
pose slices, ~50k samples, subject-grouped LOSO) to replace the rule-based
detector with a retrainable model that ports cleanly to T-maze. Full
7-phase plan in
docs/window_classifier_plan.md. - Anxiety analysis (2026-05-10) — thesis re-framed: PCA + LOOCV
classifier (Control vs Treated) and PC1-axis regression on n=29 (5–8 per
group). Empirical finding: consistent rearing-direction signal (top
feature
rear_early_fracacross all models;rear_countKruskal η²=0.20) but no result reaches statistical significance at this sample size. Power analysis: ≥20 animals/group would be needed to detect the observed effect sizes at 80% power. Thesis primary claim therefore pivots from "we showed a biological effect" to "we built a reproducible end-to-end behavioural-analysis pipeline", with empirical findings reported as a use-case demonstration with proper effect-size + CI framing. Full defensive write-up indocs/anxiety_findings_report.md.
Looking for a description of what the pipeline produces? See
docs/OUTPUTS.md— a full walkthrough of every file that lands indata/DLCfiltered/<subject>/.
We ask a single question: how do common dietary additives change rat
behaviour in open-field and T-maze assays? Twelve rats are split into four
treatment groups of three, and each rat is recorded in both arenas — giving
12 open-field videos and 12 T-maze videos (24 .avi recordings total).
| Cohort | Rats | Treatment | Role |
|---|---|---|---|
| MA1 | MA1_1–3 | none (vehicle) | control baseline |
| MA3 | MA3_1–3 | aspartame | sweetener arm |
| MA5 | MA5_1–3 | grapefruit only | grapefruit-only arm |
| MA7 | MA7_1–3 | aspartame + grapefruit | combined / interaction arm |
The grapefruit arm lets us separate a grapefruit-only effect from the aspartame + grapefruit combination, and compare both against the aspartame and control arms.
| Arena | Videos | Purpose |
|---|---|---|
| Open field | 12 | Exploration, anxiety, locomotion, rearing, grooming |
| T-maze | 12 | Decision-making, turn bias, path efficiency |
Each rat contributes one open-field recording and one T-maze recording.
Subjects are named OpenField<COHORT>_<RUN> (e.g. OpenFieldMA5_2) and
TMaze<COHORT>_<RUN>. All recordings are at 30 fps; frame f
corresponds to time f / 30 s.
flowchart TD
V["Raw videos<br/>24 × .avi<br/>12 open-field + 12 T-maze<br/>4 treatment groups × 3 rats"] --> P["Video preprocessing<br/>src/video_preprocessing.py<br/>inventory · quality QC · metadata"]
P --> D1["DLC — OFT project<br/>9 keypoints<br/>src/dlc/*"]
P --> D2["DLC — T-maze project (planned)<br/>5 keypoints: nose, ear_L, ear_R,<br/>mid_back, tail_base<br/>src/dlc/tmaze/* (planned)"]
D1 --> C1["Filtered DLC CSV (OFT)<br/>9 body parts × (x, y, likelihood)<br/>data/DLCfiltered/<group>/OpenField<...>/"]
D2 --> C2["Filtered DLC CSV (T-maze)<br/>5 body parts × (x, y, likelihood)<br/>data/DLCfiltered/<group>/TMaze<...>/"]
C1 --> B["Behaviour detection<br/>src/behavior_detection.py<br/>rule-based rearing & grooming"]
C1 --> S["Spatial analysis (OFT)<br/>analysis/open_field/run_analysis.py<br/>orbits · thigmotaxis · heatmaps"]
C1 --> M["OFT metrics<br/>analysis/open_field/oft_metrics.py<br/>locomotion · thigmotaxis · freeze · spatial entropy"]
C2 -.-> B
C2 -.-> ST["T-maze metrics (planned)<br/>turn bias · path eff. · zone dwell"]
B --> O1["*_behavior_timeline.png<br/>*_behavior_bouts.csv<br/>*_behavior_frames.csv"]
S --> O2["*_orbit_grid.png<br/>*_thigmotaxis.png<br/>*_heatmap_kde.png<br/>*_heatmap_histogram.png<br/>*_bodypart_heatmaps.png"]
M --> O3["*_oft_metrics.csv<br/>*_speed.csv · *_speed.png"]
O1 --> AX["Anxiety profile<br/>src/anxiety/profile.py<br/>OFT + bouts → 18 anxiety features"]
O3 --> AX
AX --> SR["Spatial rearing<br/>src/anxiety/spatial_rearing.py<br/>center vs wall rearing bouts"]
AX --> CL["Anxiety classifier<br/>src/anxiety/classifier.py<br/>LOOCV · LR / RF / SVM<br/>Control vs Treated (n=29)"]
AX --> CI["Composite index + PC1 regression<br/>src/anxiety/{composite_index,regression}.py"]
CL --> PR["End-to-end inference<br/>scripts/predict_anxiety_v2.py<br/>DLC CSV → per-subject prediction"]
SR --> PR
PR --> RP["reports/anxiety_predictions_v2/*_overview.png<br/>+ report.txt + report.json<br/>models/anxiety_classifier/*.pkl"]
ST -.-> R["Cross-cohort thesis report<br/>docs/final_report.md<br/>(OFT-complete · T-maze planned)"]
RP --> R
sequenceDiagram
participant User
participant Pre as video_preprocessing.py
participant DLC as src/dlc/*
participant Beh as behavior_detection.py
participant Spa as analysis/run_analysis.py
User->>Pre: 1. Inventory + QC raw videos
Pre-->>User: data/metadata.xlsx, quality_report.xlsx
User->>DLC: 2. Setup → train → infer
DLC-->>User: Filtered CSVs in data/DLCfiltered/
par Behaviour branch
User->>Beh: 3a. Detect rearing / grooming
Beh-->>User: timeline PNG + bouts/frames CSVs
and Spatial branch
User->>Spa: 3b. Orbits + heatmaps + thigmotaxis
Spa-->>User: 5 PNG visualizations
end
User->>User: 4. Compare subjects & cohorts
All screenshots below are from OpenFieldMA1_2. Every subject in
data/DLCfiltered/ has the same set of files — see docs/OUTPUTS.md
for the full per-file breakdown.
Rule-based detection of rearing (red) and grooming (green) vs. manually validated ground truth. Three rows: ground truth, detected bouts, per-frame labels.
Kernel-density estimate of the body-centre over the session. Hot spots reveal wall-hugging (thigmotaxis) vs. centre exploration.
Body-centre trajectory with the inner zone drawn at 20% margin. Used to compute the % of time spent away from the walls — a classic anxiety-like index in open-field assays.
One trajectory panel per tracked body part — useful both qualitatively (exploration pattern) and as a tracking-quality check.
End-to-end inference output from scripts/predict_anxiety_v2.py: pose
summary, OFT metrics, spatial-rearing breakdown, and Control-vs-Treated
prediction with the top driving features. One PNG per subject under
reports/anxiety_predictions_v2/.
LOOCV across n=29 (5–8 per group), rear-only feature set, logistic
regression. AUC ≈ 0.73 on held-out folds; see
docs/anxiety_findings_report.md for
the full power analysis and effect-size discussion.
Per-bout classification of rearing as center- or wall-adjacent. The
rear_center_frac axis is the most consistent group-separating signal
across all model families.
PC1 captures the rearing-direction + thigmotaxis axis; Control and Treated subjects pull in opposite directions but with substantial overlap at this sample size.
See
docs/REPO_MAP.mdfor the active vs. archived breakdown — what's on the anxiety hot path, what's planned/parallel, and what was moved toarchive/.
RatBehaviorAnalyze/
├── README.md ← this file
├── docs/REPO_MAP.md Active / planned / archived map
│
├── src/ Main pipeline code
│ ├── behavior_detection.py Rule-based rearing/grooming detector
│ ├── video_preprocessing.py Stage 01: inventory, QC, metadata
│ ├── synthesize_keypoints.py DLC keypoint synthesis / interpolation
│ ├── anxiety/ Anxiety analysis (active ML stage)
│ │ ├── config.py ARENA / INNER_ZONE / FPS single-source constants
│ │ ├── profile.py OFT + bouts → 18-feature anxiety matrix
│ │ ├── classifier.py LOOCV — LR / RF / SVM (Control vs Treated, n=29)
│ │ ├── spatial_rearing.py Center- vs wall-adjacent rearing classification
│ │ ├── composite_index.py Composite anxiety score (negative finding — see report)
│ │ └── regression.py PC1-axis regression (age/sex covariates)
│ ├── window_classifier/ Stage 06.2 — pose-window behavior classifier (planned/parallel)
│ │ ├── features.py window-level feature extraction (1 s + 3 s context)
│ │ ├── label_join.py join window features with rule-based frame labels
│ │ ├── train.py GroupKFold + LOGOCV trainer (RF / XGBoost / LightGBM)
│ │ └── infer.py DLC CSV-in → frame predictions + bout list
│ ├── requirements.txt
│ └── dlc/ DeepLabCut wrappers (OFT, 9 keypoints)
│ ├── dlc_setup.py Create project, extract & label frames
│ ├── dlc_train.py Train and evaluate the DLC model
│ └── dlc_inference.py Run inference, export filtered CSVs
│
├── analysis/ Spatial / locomotion analysis
│ ├── README.md Tool-level docs
│ ├── common/ Arena-agnostic analyses
│ │ └── speed_analysis.py Speed pipeline + cohort summary
│ ├── open_field/ OFT-specific analyses
│ │ ├── run_analysis.py One-shot runner for all 4 OFT outputs
│ │ ├── run_kare_batch.py Batch runner across all OFT subjects
│ │ ├── oft_metrics.py Per-subject OFT metric extraction (imported by anxiety pipeline)
│ │ ├── kutu_validation.py Cross-validation vs. professor's HSV-blob pipeline
│ │ ├── behavior_analysis.py Aggregated behaviour bouts → group stats
│ │ ├── cohort_stats.py Kruskal-Wallis + Dunn + PERMANOVA across cohorts
│ │ ├── label_bouts.py (planned) Interactive bout-labelling tool
│ │ ├── orbit_plot.py Per-bodypart trajectories
│ │ ├── activity_heatmap.py Body-centre KDE + histogram
│ │ ├── bodypart_heatmaps.py Per-bodypart density grid
│ │ └── show_frame_coords.py Arena-boundary picker (interactive)
│ └── tmaze/ T-maze-specific analyses (parallel layout, planned)
│
├── scripts/ Entry-point CLIs
│ ├── predict_anxiety_v2.py End-to-end anxiety inference (DLC CSV → prediction)
│ ├── predict_anxiety_v3_gui.py Streamlit GUI wrapper around v2 (browser-based)
│ ├── debug_features.py Quick feature/threshold diagnostic
│ └── legacy/ Archived early prototypes (Windows-path hardcoded)
│
├── tools/
│ └── convert_labeled_to_video_format.py
│
├── data/
│ ├── DLCfiltered/ One folder per cohort/subject (inputs + outputs)
│ │ ├── control/ MA1 cohort
│ │ │ └── OpenFieldMA1_<RUN>/ ← see docs/OUTPUTS.md for the per-subject file set
│ │ ├── ASP/ MA3 cohort
│ │ ├── Greyfurt/ MA5 cohort
│ │ ├── ASP ve Greyfurt/ MA7 cohort
│ │ └── Kare/ Professor's _res.mat reference (Kutu_v1, OFT only)
│ ├── anxiety_features.csv Cross-subject 18-feature anxiety matrix
│ ├── metadata.xlsx Per-video metadata
│ ├── oft_metrics_all.csv Cross-subject OFT metric table
│ ├── behavior_summary.csv Cross-subject rearing/grooming summary
│ ├── kutu_validation_summary.csv DLC vs. HSV-blob speed cross-check
│ ├── speed_summary_all.csv Mean/instantaneous speed per subject
│ ├── part0_trajectories/ Quick-look trajectory PNGs
│ ├── quality_plots/ Per-video brightness plots
│ └── quality_report.xlsx QC summary
│
├── models/
│ ├── anxiety_classifier/ {lr, rf, svm}_rearonly.pkl + scaler (Control vs Treated)
│ ├── anxiety_regression/ PC1-axis regression models
│ └── window_classifier/ Pose-window classifier pickles (planned/parallel)
│
├── reports/
│ ├── anxiety_predictions_v2/ Per-subject v2 inference: overview.png + report.{txt,json}
│ ├── anxiety_classifier_*.csv Metrics / importance / predictions (rear-only + full)
│ ├── anxiety_regression_*.csv PC1 regression metrics + predictions
│ ├── anxiety_pca.png Feature-space PCA across all subjects
│ ├── anxiety_boxplots.png Per-feature group distributions
│ ├── spatial_rearing_stats.csv Center vs wall rearing summary
│ ├── composite_anxiety_*.csv Composite index (negative finding)
│ ├── cohort_kruskal_wallis.csv Per-feature KW (H, p_perm, ε², q_bh)
│ ├── cohort_dunn_posthoc.csv Pairwise Dunn z + p (within-feature BH)
│ ├── cohort_permanova.csv Multivariate pseudo-F + R² + p_perm
│ ├── window_classifier_*.csv Window-level model metrics (planned/parallel)
│ └── figures/ Confusion matrices, ROC, SHAP, spatial rearing PNGs
│
├── archive/ Superseded code — kept for history, not on hot path
│ ├── README.md Why each thing is here
│ ├── scripts/predict_anxiety.py v1 (replaced by predict_anxiety_v2.py)
│ ├── src/ {train_baseline, features, _metrics_minimal,
│ │ visualize_reports, train_anxiety_demo}.py
│ └── models/classifier/ Old 4-target baseline pickles (25 files)
│
└── docs/ All documentation (single root)
├── REPO_MAP.md Active / planned / archived map
├── OUTPUTS.md Guide to every output file under data/DLCfiltered/
├── WORKFLOW_SUMMARY.md Filtering pipeline & parameters (analysis-side)
├── final_report.md OFT-complete thesis chapter + inference roadmap
├── anxiety_findings_report.md Anxiety analysis — pipeline pivot + empirical findings
├── anxiety_progress_2026-05-10.md Latest anxiety progress log
├── window_classifier_plan.md Window-level classifier — 7-phase implementation plan
├── tmaze_keypoints_and_layout.md T-maze keypoint plan + repo layout (5-point DLC)
├── behavior_detection_documentation.md Detector design (algorithmic)
├── behavior_detection_documentation.legacy.md Detector design rationale (Turkish narrative)
├── behavior_comparison.md Cohort-vs-cohort behavioural comparison
├── yapilanlar.md Running progress log (Turkish, historical)
├── DEEPLABCUT_PIPELINE.md DLC workflow (10 stages)
├── RAT_TMAZE_PROJECT.md Earlier T-maze roadmap (superseded — see tmaze_keypoints_and_layout.md)
└── plans/ Per-stage design docs (STAGE_01 … STAGE_07, historical)
conda create -n dlc python=3.10 -y
conda activate dlc
pip install -r src/requirements.txtRequires CUDA 11.8 or 12.1 with an NVIDIA GPU for DLC training (tested on RTX 3060 6 GB). Behaviour detection and spatial analysis run on CPU.
Assuming a DLC-filtered CSV is already in place at
data/DLCfiltered/OpenFieldMA5_2/OpenFieldMA5_2.csv:
# 1. Detect rearing & grooming
python src/behavior_detection.py \
--csv data/DLCfiltered/OpenFieldMA5_2/OpenFieldMA5_2.csv
# 2. Pick arena boundaries interactively (first subject only)
python analysis/open_field/show_frame_coords.py \
--video data/DLCfiltered/OpenFieldMA5_2.mp4
# 3. Generate all 4 spatial visualizations in one shot
python analysis/open_field/run_analysis.py --arena 396 776 153 530
# 4. Run end-to-end anxiety inference (DLC CSV → Control/Treated prediction)
python scripts/predict_anxiety_v2.py \
--csv data/DLCfiltered/OpenFieldMA5_2/OpenFieldMA5_2.csv
# 5. (alternative) Launch the v3 GUI — upload a DLC CSV in your browser
pip install streamlit # one-time
streamlit run scripts/predict_anxiety_v3_gui.pyOFT outputs land next to the input CSV. Anxiety v2 outputs land in
reports/anxiety_predictions_v2/. The v3 GUI wraps the same v2 pipeline,
shows the overview PNG + Turkish report inline, and offers download
buttons — optionally also writes to reports/anxiety_predictions_v2/.
| Stage | Script / folder | What it does | Status |
|---|---|---|---|
| 01 | src/video_preprocessing.py |
Video inventory, QC (brightness, FPS), metadata export | done |
| 02 | src/dlc/dlc_setup.py |
Create DLC project (OFT, 9 keypoints), extract frames, launch labelling GUI | done |
| 02 | src/dlc/dlc_train.py |
Train and evaluate the OFT DLC model | done |
| 02 | src/dlc/dlc_inference.py |
Run inference, export filtered tracking CSVs | done |
| 02B | src/behavior_detection.py |
Rule-based rearing / grooming classifier (OFT) | done |
| 03A | analysis/open_field/run_analysis.py |
Orbits, thigmotaxis, KDE & per-bodypart heatmaps (OFT) | done |
| 03A.1 | analysis/common/speed_analysis.py |
Speed pipeline + cohort summary | done |
| 03A.2 | analysis/open_field/kutu_validation.py |
Cross-validation vs. professor's 2017 HSV-blob MATLAB pipeline | done |
| 03A.3 | analysis/open_field/oft_metrics.py |
Per-subject OFT metric extraction (locomotion, thigmotaxis, freeze, spatial entropy) | done |
| 05 | src/anxiety/profile.py |
Anxiety feature engineering — 18 metrics per subject (src/features.py archived) |
done |
| 06 | src/anxiety/classifier.py |
LOOCV — LR / RF / SVM, Control vs Treated (n=29). Old 4-target baseline src/train_baseline.py archived |
done |
| 06 | src/anxiety/spatial_rearing.py |
Center- vs wall-adjacent rearing — primary group-separating signal | done |
| 06 | src/anxiety/{composite_index,regression}.py |
Composite anxiety index + PC1-axis regression | done |
| 07 | scripts/predict_anxiety_v2.py + docs/anxiety_findings_report.md |
End-to-end inference (DLC CSV → prediction) + thesis write-up | done |
| 07.GUI | scripts/predict_anxiety_v3_gui.py |
Streamlit GUI wrapper — browser upload → overview PNG + TR report + downloads | done |
| 06.1 | analysis/open_field/cohort_stats.py |
Non-parametric cohort effect — Kruskal-Wallis + Dunn + PERMANOVA (replaces failed cohort classifier) | done |
| 06.2 | src/window_classifier/ + docs/window_classifier_plan.md |
Window-level behaviour classifier — implementation parallel, not yet integrated into anxiety pipeline | planned |
| — | docs/tmaze_keypoints_and_layout.md |
T-maze keypoint plan + repo layout (5-point DLC project, separate from OFT) | planned |
| 02-T | src/dlc/tmaze/* (planned) |
T-maze DLC project — separate model, 5 keypoints | planned |
| 02B-T | src/behavior_detection.py (refactor) |
Make detector keypoint-profile aware so the same code serves both arenas | planned |
| 03B | src/tmaze/tmaze_metrics.py (planned) |
Turn bias, path efficiency, zone dwell, decision latency | planned |
| 04 | analysis/tmaze_path_plot.py (planned) |
Path overlay + zone heatmap | planned |
Design docs for each stage live in docs/plans/. T-maze keypoint and
layout decisions are consolidated in
docs/tmaze_keypoints_and_layout.md.
Rule-based detector in src/behavior_detection.py. No ML training needed —
thresholds on postural features separate behaviours.
flowchart LR
F["DLC features per frame<br/>htdist, fp_hp_vert, nose2fp,<br/>nose_x/y, body_vel"] --> G["Grooming posture?<br/>nose near forepaws<br/>+ low velocity<br/>+ nose inside arena"]
G -- yes --> RC["Compact-rearing rule<br/>excludes grooming posture"]
G -- no --> RC
RC --> R["Rearing rules (any fires)<br/>R1 compact · R2 top-wall<br/>R3/4 bottom-wall · R5 side-wall<br/>R6 wall-press"]
R --> L["Per-frame label<br/>rearing > grooming > other"]
L --> B["Bout grouping<br/>merge gap ≤ 15 frames<br/>discard < 10 frames"]
Full rationale (thresholds, failure modes, ground-truth validation) in
docs/behavior_detection_documentation.md.
Hand-labelled windows for a subset of subjects are stored in
GROUND_TRUTH_BY_SUBJECT in src/behavior_detection.py and rendered as
the top row of the timeline plot. Current coverage:
| Subject | Rearing GT | Grooming GT |
|---|---|---|
| OpenFieldMA1_2 | 9 windows | 2 windows |
| OpenFieldMA5_1 | — | 5 windows |
| all others | not yet entered | not yet entered |
| Metric | Psychological relevance |
|---|---|
| Rearing bouts & total rearing time | Exploration / vigilance |
| Grooming bouts & total grooming time | Stress recovery / displacement |
| Thigmotaxis (% time in inner zone) | Anxiety-like behaviour |
| Total distance travelled | Locomotor activity |
| Mean velocity | Arousal / inhibition |
| Spatial KDE hot-spot count | Exploration pattern |
DLC keypoints: nose, ear_L, ear_R, mid_back, tail_base. Front paws and
hip/shoulder are excluded — top-down camera + small subjects + corridor
wall occlusion make them more noise than signal. Full rationale and
per-behaviour metric mapping in
docs/tmaze_keypoints_and_layout.md.
| Metric | Psychological relevance | Keypoints used |
|---|---|---|
| Turn bias (L/R ratio) | Lateralization, perseveration | nose, mid_back, tail_base |
| Decision latency | Anxiety, deliberation | mid_back, tail_base |
| Path efficiency | Cognitive load | mid_back, tail_base |
| Zone dwell time | Preference / avoidance | mid_back |
| Backtrack rate | Memory deficit, uncertainty | (nose − mid_back) heading |
| Rearing (top-down) | Exploration / vigilance | nose, mid_back, tail_base |
| Grooming (top-down) | Stress recovery / displacement | nose, ear_L, ear_R, mid_back |
| Component | Tool |
|---|---|
| Pose estimation | DeepLabCut (PyTorch backend) |
| Video processing | OpenCV |
| Data processing | Pandas, NumPy, SciPy |
| Visualization | Matplotlib, Seaborn |
| ML models (planned) | Scikit-learn, XGBoost, PyTorch |
| Notebooks | Jupyter Lab |
docs/REPO_MAP.md— active vs. planned vs. archived map — read this first to orientdocs/anxiety_report.md— CONSOLIDATED anxiety report (2026-05-11) — pipeline · findings · model decision logic · config-mismatch incident · reproducibility · thesis cümleleri. Tek otoritatif kaynak.docs/OUTPUTS.md— what every file indata/DLCfiltered/<subject>/means and how to read itdocs/final_report.md— consolidated OFT thesis chapter + ML inference roadmapdocs/anxiety_findings_report.md— (historical 2026-05-10) anxiety pipeline pivot + first-pass findings — superseded byanxiety_report.mddocs/anxiety_progress_2026-05-10.md— (historical 2026-05-10) spatial rearing + v2 inference progress log — superseded byanxiety_report.mddocs/tmaze_keypoints_and_layout.md— T-maze keypoint plan + repo layout (5-point DLC project)docs/window_classifier_plan.md— window-level behaviour classifier — 7-phase implementation plandocs/behavior_detection_documentation.md— behaviour-detector design (algorithmic)docs/behavior_comparison.md— cohort-vs-cohort behavioural comparisondocs/yapilanlar.md— running progress log (Turkish)analysis/README.md— spatial-analysis tools (quickstart + parameters)docs/WORKFLOW_SUMMARY.md— filtering pipeline detailsdocs/DEEPLABCUT_PIPELINE.md— full DLC workflow (10 stages)docs/RAT_TMAZE_PROJECT.md— earlier roadmap (single-DLC-model assumption superseded bydocs/tmaze_keypoints_and_layout.md)docs/behavior_detection_documentation.legacy.md— behaviour-detector design rationale (Turkish)docs/plans/— per-stage design documents
| Stage | Status |
|---|---|
| 01 — Video preprocessing / QC | done |
| 02 — DeepLabCut pose estimation (9-keypoint OFT model) | done |
| 02B — Rule-based behaviour detection (rearing / grooming) | done, validated on MA1_2 / MA5_1 |
| 03A — Spatial analysis (orbits, thigmotaxis, KDE, per-bodypart) | done for MA1 / MA3 / MA5 / MA7 |
| 03A.1 — Speed pipeline + cohort summary | done |
| 03A.2 — Cross-validation vs. professor's HSV-blob pipeline | done (~14% mean-speed bias, uniform across cohorts) |
05 — Anxiety feature engineering (src/anxiety/profile.py, 18 metrics) |
done — src/features.py (old 20+ metrics) archived under archive/src/ |
06 — Anxiety classifier (src/anxiety/classifier.py, LOOCV LR/RF/SVM, Control vs Treated n=29) |
done — AUC≈0.73 rear-only; old 4-target baseline archived |
06 — Spatial rearing (src/anxiety/spatial_rearing.py) |
done — rear_center_frac is the primary group-separating signal |
06 — Composite index + PC1 regression (src/anxiety/{composite_index,regression}.py) |
done — composite reported as negative finding |
06.1 — Cohort statistics (analysis/open_field/cohort_stats.py, KW + Dunn + PERMANOVA) |
done — center_zone_entries survives within-feature FDR |
06.2 — Window-level behaviour classifier (src/window_classifier/) |
implemented, parallel — not yet integrated into anxiety pipeline |
07 — End-to-end inference (scripts/predict_anxiety_v2.py) + thesis write-up (docs/anxiety_findings_report.md) |
done |
| Stage | Status |
|---|---|
Keypoint plan + file-layout decision (docs/tmaze_keypoints_and_layout.md) |
done (2026-05-05) |
| 02-T — DLC project setup (5 keypoints, separate from OFT) | planned |
| 02-T — DLC labelling + training | planned |
| 02B-T — Behaviour detection refactor (keypoint-profile aware) | planned |
| 03B — T-maze metrics: turn bias, path efficiency, zone dwell, decision latency | planned |
| 04 — Path overlay + zone heatmap | planned |
Integration into docs/final_report.md (T-maze chapter) |
planned |









