henhaobio
Sign in

bioRxivNeurosciencePreprint

ABEL: an active-learning behavior estimation and labeling platform

Jobe L. RitchieiD, Brianna A. GeorgeiD, Alison V. RolandiD, Caroline G KriemaniD, Brooke N. BenderiD, Maya R EberleiD, Grace A StysiD, William C. DeriD, Lili S KooymaniD, Adriane M. LawesiD, Rajon T. ScottiD, Todd K. O’BuckleyiD, Marissa M. R. McLeaniD, Chloe J. GallagheriD, Joyce BesheeriD, Thomas L. KashiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Detailed behavior analysis is essential for thorough characterization of ethologically relevant behaviors in model organisms, yet manual annotation of the full behavioral repertoire remains subjective, time intensive, and susceptible to observer error. Advances in machine learning have enabled high-throughput pose estimation on recorded video, but tools for behavior classification from pose and video data are still developing. Instead of hand-scoring every frame of video, experimenters can instead label a small subset of video frames and software trained through machine learning makes predictions on the rest. Here, we present an Active-learning Behavior Estimation and Labeling (ABEL) platform that uses clip-level active learning (i.e., human labeling of short video snippets) with multimodal features (pose, video, context/ROI) to train robust behavior classifiers. We rigorously validated ABEL-derived behavior predictions against expert human observers and field-standard automated software, across diverse rodent behavioral assays. Across eight assays and 45 behaviors, model training required 19.5 hours of human annotation in total, with the reviewer scoring ~8% of available video. Models trained in ABEL achieved a mean precision-recall area under the curve (PR-AUC - a 0-1 score of how well a model balances missed detections against false alarms, with 1 being perfect) of 0.90 (SD 0.09, range 0.60-0.99), with no association between performance and behavior prevalence (r = 0.20). This was aided by custom tools, Essence Extractor and UMAP Interactive Selection, for targeted discovery of high probability clips which reduce the clip review needed to find a rare behavior 6-fold relative to random sampling and 10-fold relative to labeling whole videos. As a biological validation, we assessed how ABEL-derived behaviors relate to underlying neuronal calcium dynamics. Behavior labels were tightly synced with neuronal signatures distinct from ambiguous behavior and randomly chosen, behavior-unrelated time windows (shuffle control). Together, these data indicate that ABEL provides an efficient platform for frame-precise classification of distinct ethologically relevant behaviors.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (9 Sept 2026) · CC BY 4.0

A no-code clip-level active-learning platform combining pose, video and region-of-interest features trains rodent behavior classifiers (mean precision-recall AUC 0.90, 45 behaviors, eight assays) largely agreeing with expert human scoring.

Pose tracking with DeepLabCut or SLEAP is mature, but turning keypoints into named behaviors still relies on unsupervised motif discovery (MoSeq), whose motifs may not map onto ethological behaviors, or supervised labeling (SimBA) that is costly in annotator time. ABEL moves the frame-level active learning of A-SOiD to short clips, adds video and region-of-interest features, and provides a graphical interface with rare-behavior mining tools that needs no coding.

Key results

  1. ShownAcross the assays described, ABEL behavior measures agreed with expert human scoring. On a previously published home-cage dataset, ABEL matched or exceeded SimBA in concordance.

    Home cage (n = 10 mice), Lin's concordance correlation coefficient (CCC), ABEL vs SimBA: rearing 0.92 vs 0.81, grooming 0.97 vs 0.98, digging 0.98 vs 0.87. Novel object CCC 0.91 and latency to eat CCC 0.89 (n = 10 mice each); rat predator-odor immobility CCC 0.94 (n = 7 rats). Elevated plus maze protected-dip ratio: ABEL CCC 0.76 (bias -10.99%) vs 0.90-0.95 for three humans. · ABEL agrees with expert human observation across…

  2. ShownAcross eight assays and 45 behaviors, models reached a high precision-recall area under the curve (PR-AUC) after about 8% of the available video was labeled. Performance showed no detected link with behavior prevalence.

    19.5 h of annotation in total; mean PR-AUC 0.90 (SD 0.09, range 0.60-0.99); no detected association with prevalence of 0.26-44.39% of reviewed segments (Pearson r = 0.20, 95% CI -0.10 to 0.46, p = 0.19). Mean Cohen's kappa vs expert annotation 0.79; 96% of behaviors above 0.6. Clip-level metrics at a 250-clip budget: mean F1 0.70, PR-AUC 0.80. · Abstract; ABEL is data-efficient and provides advanced…; Statistical Analysis

  3. ShownVideo-embedding and region-of-interest features add information beyond pose. Removing them lowers accuracy for most behaviors, and they help separate behaviors whose poses look alike.

    Withholding video features lowered held-out F1 in 35 of 45 behaviors (paired t-tests, Benjamini-Hochberg q < 0.05). ROI features raised held-out F1 by a mean of 0.13 in projects with defined ROIs (28 of 32 behaviors improved; data not shown). Context and/or video features improved discrimination in 70 of 82 behavior pairs (median 69.3% reduction in remaining pose-only error). · ABEL is data-efficient and provides advanced…; Statistical Analysis

  4. SuggestedThe best clip-surfacing strategy depends on the clip budget: Essence Extractor finds positives fastest early and active learning does better later. Targeted strategies beat random clips and whole-video scanning for finding rare behaviors.

    At 25 clips, Essence Extractor surfaced 1.89x more positives than UMAP Interactive Selection and 1.55x more than active learning; at 300, active learning led (1.36x and 1.71x, respectively). The targeted tools found ~5-6x more positives than random selection. For the eight rarest behaviors (1.23 ± 0.41% SEM prevalence), Essence Extractor and active learning were similar (p = 0.76); effort assumed 4 s per clip. · Abstract; ABEL is data-efficient and provides advanced…; Active-learning machine learning; Statistical Analysis

  5. SuggestedIn the novelty suppressed feeding (NSF) assay, bouts labeled by ABEL carried distinct calcium signatures in the bed nucleus of the stria terminalis (BNST) and the central nucleus of the amygdala (CeA). Ambiguous periods looked like random windows.

    Fiber photometry in 7 male mice; functional linear mixed models: approach, sniff, groom, eat and rear differed from pseudorandom windows in both regions (all p < 0.01); ambiguous vs random p > 0.05. CeA significance windows for rear and approach began within 1-2 frames (33-67 ms at 30 Hz) of bout start; eat windows preceded eating by about 30-40 frames. NSF models F1 = 0.90 ± 0.03. · ABEL extracts neurophysiologically distinct…; Functional linear mixed model analysis; Figure legends

Methods, data and reproducibility

Pose-tracked videos are cut into 0.5 s clips. For each clip, about 2080 pose, R3D-18 video-embedding and region-of-interest (ROI) context features feed one-vs-rest XGBoost classifiers trained in a clip-level active-learning annotation loop, then applied densely with mutual inhibition, thresholds and bout filtering. Outputs were compared with human scorers, SimBA, EthoVision XT and photometry.

  • Previously published overhead home-cage videos (rearing, grooming, digging), also scored by SimBA and an expert human · mouse · n = 10 mice
  • Elevated plus maze videos; head dips and protected head dips scored by three human observers · mouse · n = 10 mice
  • 3-chamber social assay videos, compared with an expert human and EthoVision XT zone time · mouse · n = 8 mice
  • Novel object recognition videos (novel and familiar object investigation) · mouse · n = 10 mice
  • Novelty suppressed feeding videos (approach count, latency to eat) · mouse · n = 10 mice
  • Predator odor (TMT) assay, side-view videos (digging, immobility) · rat · n = 7 rats
  • NSF assay with dual-site fiber photometry (AAV8-hSyn-GCaMP8f in BNST and CeA) · mouse · n = 7 males
  • All annotated projects pooled for model benchmarking (C57BL6/J mice; Wistar and Long-Evans rats) · mouse; rat · n = 8 assays, 45 behaviors
Methods & tools
ABEL v0.11.0, DeepLabCut (DEKR_w18 backbone), SLEAP, XGBoost 3.2.0, scikit-learn 1.7.2, R3D-18, UMAP, SimBA, EthoVision XT (Noldus), TRACY Photometry Processing Suite, fastFMM 1.0.1, fiber photometry (AAV8-hSyn-GCaMP8f)
Software & environment
ABEL v0.11.0, run locally on Windows; xgboost 3.2.0, scikit-learn 1.7.2; Python 3.11 with NumPy 1.26, SciPy 1.11, pandas 2.1; fastFMM 1.0.1 in R 4.6.0; GraphPad Prism 11.01; DeepLabCut DEKR_w18 models trained for 200 epochs.
Assay conditions & compute
ASUS laptop with Intel Core i9-13980HX CPU, 64GB memory and NVIDIA 4070 laptop GPU; XGBoost trained on CUDA with a slower CPU fallback. Human labeling: 19.5 hours in total, about 2-5 hours per assay.
Reproducibility gaps
Data available only from the corresponding author on reasonable request; ABEL is free for academic use (terms for other users not stated); the full feature list (Supplemental Table 1) and assay protocols are in supplements not included in this text.
Limitations (6)
  • Because ABEL is supervised, it cannot find behaviors the user did not anticipate. Parallel unsupervised tools (UMAP selection, motif discovery) only partly offset this, and the labeling burden stays with the scorer. · stated by the authors
  • Training labels came from a single annotator, which may limit the validity and generalizability of the models, although comparisons with other expert observers showed high reliability. · stated by the authors
  • All models used a fixed 0.5 s segmentation window, which may miss features that unfold over longer time scales. · stated by the authors
  • Labeling-effort comparisons rest on assumed times (4 s per clip, 1x real time for whole videos) that may favor ABEL; logged review timestamps showed a median of 1 s per clip decision (43431 decisions). · stated by the authors
  • The neural validation covered only two brain structures in a small sample, and eating-specific calcium signal could not be separated from the food-directed behaviors that precede eating. Other circuits or behaviors may align less tightly. · stated by the authors
  • Rare-behavior effort gains are reported inconsistently: abstract 6-fold vs random and 10-fold vs whole videos, Introduction 1.8x, Discussion 1.85x vs active learning, yet Results find Essence Extractor and active learning similar (p=0.76); Fig. 3I legend says 50 positive clips, the text 25. · noted in this summary

Read this if you score rodent behavior from DeepLabCut or SLEAP tracking and want a no-code supervised classifier that limits annotation effort, especially for rare behaviors or for alignment with fiber photometry.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISYttri '26Harris '23Li '22Kozlova '25Gabriel '22Bohnslav '21
73% semantic matchNeuroscience
Eric A. Yttri, Aden Eagle, Brayden Kohler+2 more·bioRxiv·Oct 20, 2026