Machine Learning to Analyze Alternating Treatments Graphs
A deep-learning tool can reliably spot differentiated ATD effects, giving you a quick, unbiased second opinion on your visual analysis.
01Research in Context
What this study did
Kausch et al. (2026) built a deep-learning model. They fed it thousands of alternating-treatments graphs. The goal was to see if a computer could spot when one line clearly beats the other.
The team trained the network on graphs where experts had already marked “differentiated” or “not differentiated.” Then they tested the model on new graphs it had never seen.
What they found
The model got over 90 % of the calls right. It rarely cried wolf; false alarms stayed low.
The machine matched expert eyes, but it did the job in seconds and never got tired.
How this fits with other research
Kahng et al. (2010) showed that well-trained BCBAs already agree strongly when they eyeball graphs. Kausch keeps that high agreement, yet swaps human eyes for silicon ones.
Wolfe et al. (2016) found that experts barely agree on multiple-baseline graphs. Kausch gives the field an objective referee that could settle close calls.
Dowdy et al. (2024) warned that small tweaks like axis ratio can fool human viewers. Kausch’s model ignores those cosmetic tricks; it learns the data pattern itself.
Why it matters
You still make the final call, but now you can run the graph through an ML screen before you sign off. If the model flags “differentiated” and you see the same, your confidence grows. If it disagrees, you know to look again. One click gives you a second opinion that is fast, free of bias, and easy to document for peer review or insurance.
Want CEUs on This Topic?
The ABA Clubhouse has 60+ free CEUs — live every Wednesday. Ethics, supervision & clinical topics.
Join Free →Upload your next ATD graph to the free web demo and compare the model’s call to your own visual inspection.
02At a glance
03Original abstract
Alternating treatment designs (ATD) allow practitioners to rapidly compare the effectiveness of multiple interventions without needing long baseline phases. ATDs are typically evaluated using visual analysis and effect size analysis. However, past research suggests that different raters may come to different conclusions when analyzing these graphs. Furthermore, agreement between raters does not equate to accuracy. Artificial Intelligence (AI) technologies may improve the replicability of decision-making while performing adequately with smaller datasets. Within AI, Machine Learning (ML) algorithms can learn patterns and make data-driven predictions. These algorithms can quantify complex, intuitive patterns directly from data that would be difficult or impractical to define explicitly with traditional programming rules. This study investigates the use of ML technology to analyze ATD graphs. Specifically, the researchers examined which feature engineering techniques resulted in predictive performance for an ML model trained on simulated or non-simulated data and tested on non-simulated data. The best-performing models achieved classification accuracy above 90% with type 1 error rates below 15%. Deep neural networks (DNNs) led to the highest accuracy in detecting differentiated effects in ATD graphs while minimizing false positives. These results provide initial evidence that embedding DNNs into analysis of ATD single-case graphs could enhance the replicability of data analysis.
Journal of Behavioral Education, 2026 · doi:10.1007/s10864-026-09616-z