Machine learning identifies depression as key barrier to chatbot-delivered insomnia therapy adherence

A chatbot named Master Sleep was supposed to guide people through a month of cognitive behavioral therapy for insomnia, delivered entirely through text messages on Telegram. The program worked—at least for those who stuck with it. But most people didn't. Only 19 percent of the 75 participants who enrolled and met the study criteria actually completed all 28 days. The rest dropped out somewhere along the way, and researchers wanted to know why.

The study, conducted in Hong Kong between April 2023 and May 2024, recruited people through social media ads and community outreach. Participants had to be between 18 and 65, speak Cantonese, have a smartphone with internet access, and score at least 8 on the Insomnia Severity Index—a clinical threshold for actual insomnia. The chatbot delivered the standard four pillars of insomnia treatment: sleep education, behavioral strategies like stimulus control and sleep restriction, cognitive restructuring to challenge unhelpful thoughts about sleep, and stress management techniques. Each day, participants logged their sleep data, received a new lesson, and could ask questions. The chatbot used machine learning to understand their messages and respond appropriately, even detecting emotional distress when it appeared.

What made some people persist while others vanished? Researchers used machine learning models to hunt for patterns in baseline characteristics—demographics, sleep measures, depression and anxiety scores, beliefs about sleep, sleep hygiene habits. They found that depressive symptoms were the single strongest predictor of dropout. People who reported higher depression at the start were significantly less likely to complete the program. The mechanism is straightforward: depression drains motivation and makes cognitive tasks feel harder. A person already struggling with low mood and low energy faces a steeper hill when asked to engage with a digital intervention, even one designed to help them sleep better.

But the study also uncovered something unexpected about daily patterns. On nights when people slept poorly—when they took longer to fall asleep or woke up repeatedly in the middle of the night—they were more likely to use the chatbot the next day. They weren't avoiding it; they were reaching for it. This suggests that poor sleep itself can drive engagement, at least temporarily. A person lying awake at 3 a.m. might be more motivated to try the intervention than someone sleeping reasonably well. The researchers also tested whether using the chatbot close to bedtime—within two hours of sleep—would disrupt sleep that night. It didn't. Despite the common concern about blue light from screens before bed, nighttime use of the intervention showed no correlation with worse sleep. On average, participants used the chatbot only twice per night, suggesting that most heeded the advice to limit late-night engagement.

The machine learning models showed promise in identifying who would struggle with adherence overall, though predicting day-to-day adherence proved harder. The researchers used four different algorithms—logistic regression, support vector machines, random forests, and gradient boosting—and tested them rigorously with cross-validation. The models worked best when trained on baseline characteristics, suggesting that a therapist could theoretically screen new patients and flag those at high risk of dropout before they even start.

What happened to people who did complete the program? Their sleep improved. Sleep efficiency—the percentage of time in bed actually spent asleep—increased for both those with good adherence and those who only partially engaged. People with strong adherence saw a more dramatic reduction in wake time after sleep onset, the frustrating experience of waking up in the middle of the night. This hints at something important: even incomplete engagement with the intervention produced measurable benefit. Some of those who dropped out early may have done so not because the program failed them, but because their sleep improved enough that they felt they could stop.

The study is small and has real limitations. Seventy-five people is not a large sample. The researchers relied on self-reported sleep data rather than objective measures like actigraphy. They tested the chatbot only in one cultural context, in Cantonese, with a population that may not represent other regions or languages. And they excluded people with existing psychiatric diagnoses, which means the findings don't apply to the people most likely to have both insomnia and depression. Still, the core finding stands: depression is a barrier to engagement with digital mental health tools, and machine learning can spot that barrier before treatment begins. The next step is obvious—therapists could use these predictions to intervene early, offering hybrid human-AI support to people flagged as high-risk, rather than hoping they'll persist with a fully automated program alone.