Rapid and Reliable Computational Markers for Predicting Daily Smoking Behavior and Smoking Cessation Treatment Outcome

Nicotine addiction is a complex disorder shaped by factors such as craving, mood, and neurocognitive processes. While ecological momentary assessment (EMA) provides a real-time method for capturing dynamic changes in behavior, traditional tasks and surveys are often too lengthy and demanding for repeated use in clinical settings. Integrating EMA with computational approaches offers a promising solution to predict smoking behavior dynamically while addressing the practical limitations of conventional assays, paving the way for more effective and scalable interventions. To evaluate the predictive value of computational markers derived from decision-making tasks and ecological momentary assessment (EMA) data for short-term (daily smoking behavior) and long-term (cessation success) outcomes, and to assess the timing and amount of data collection needed for prediction. 79 daily smokers (mean age 25.64 years, 83% male) took part in a longitudinal experimental study involving EMA surveys of psychological states and decision-making tasks, delivered daily via a smartphone app, while undergoing a 5–6 week smoking cessation program. Using a machine-learning methodology (adaptive design optimization, ADO) to effectively generate task variables, we estimated computational markers from just 20 to 30 trials per day, reducing task length and participants burden. A time-lagged model incorporating both computational markers and self-reported daily psychological states provided the most accurate prediction of next-day smoking behavior. Higher levels of craving, depression, and ambiguity tolerance in decision-making on the previous day were significantly predictive of increased smoking amount the following day. Smoking cessation status at the end of treatment was most strongly predicted by lower discounting rates, reduced craving and stress, a longer smoking history, and greater engagement in treatment (AUC = 0.76). Notably, models based on data collected during the first week of follow-up, either on the decision-making tasks (AUC = 0.74) or psychological variables (AUC = 0.73), demonstrated comparable predictive accuracy for end-of-treatment smoking cessation. Combining computational markers with EMA data offers a dynamic and efficient approach for predicting smoking behavior and cessation success and holds promise for clinical applications.