Smartphone movement data can reliably predict smoking lapses and cravings to enable timely smoking cessation support

www.nature.com/scientificreports OPEN Smartphone movement data can reliably predict smoking lapses and cravings to enable timely smoking cessation support Maryam Abo-Tabik1,4, Nicholas Costen2 & Yael Benn3,4 Decades of research aiming to develop effective smoking interventions have identified triggers that contribute to failed quitting attempts including environmental (e.g. location), social (presence of other smokers), or internal (e.g. stress). Here, it is shown for the first time that passively collected movement data from smokers’ smartphones’ sensors (accelerometer, gyroscope and magnetometer) can be used to predict smoking-behaviour. Feeding the movement data into a Deep Learning (DL) model (1D-CNNBiLSTM), smoking-behaviour was predicted with 85% accuracy within the subsequent 5-minute window. This compares to 63% accuracy when using traditional triggers (e.g. time of the day). Crucially, movement data can be used to predict high-craving incidents and lapses in the 3 months period following quitting smoking with similarly high accuracy, even when predictions are made without any personal data (i.e. when the model is trained using only data from other smokers). These findings can transform smoking-cessation apps, enabling the provision of just-in-time personalised support to those wishing to quit smoking. Importantly, the findings have implications beyond smoking-cessation applications, by revealing that human movements, largely overlooked to date, can be used for early detection of, and intervention for, health (and other) behaviours, including those that are not genetic or typically characterised by changes in movement. Smoking is a public health emergency, increasing the risks for multiple health conditions, which result in the death of around 8 million people worldwide annually1,2. While many smokers try to give up smoking, quitting success rate remains low3. To support smokers who wish to quit, many mobile-based smoking cessation apps have been developed4. While there are several popular applications that are effective in improving quitting rate (for example, the United Kingdom National Health Service quit smoking app5), these apps do not make use of the full range of the currently available technology4, and hence provide less than optimal support. Studies indicate that providing just-in-time targeted intervention can minimise relapse incidents6,7. To enable the delivery of such timely intervention (e.g. via a mobile app), several attempts have been made to utilise Machine Learning (ML) algorithms for predicting smoking-behaviour based on known smoking riskfactors such as high urges8, geographical location9 or social triggers10. However, these studies suffer from several limitations. For one, the data they are based on largely relies on self-reporting from smokers (e.g. 8), which is considered unreliable, as self-reporting is often affected by recall issues, social bias and low commitment, leading to potentially misleading results in health outcomes and behaviour assessments11. Multiple mobile sensors such as accelerometers or Global Positioning System (GPS) are now commonly embedded within smartphones and smartwatches, and they offer a superior method for passively collecting realtime and objective data. Such data have increasingly been used within ML algorithms to analyse and understand complex behavioural patterns to improve insight into different behaviours and medical conditions12. Recent studies have tended to use GPS signals to predict smoking events. Abo-Tabik et al.13 fed passively collected GPS and accelerometer data from smartphones, through a combined 1D Convolutional Neural Network (1D-CNN) and Control Theory model. While the model achieved 0.74 accuracy, it was trained only on regular, pre-quit smoking data, meaning its ability to predict lapses after quitting remains unclear. Furthermore, the model was restricted to predicting events within a 60-minute window, limiting its utility for “just-in-time” 1Department of Computer Science, University of Central Lancashire, Preston, UK. 2Department of Computing and Mathematics, Manchester Metropolitan University, Manchester, UK. 3School of Psychology, Manchester Metropolitan University, Manchester, UK. 4These authors contributed equally: Maryam Abo-Tabik and Yael Benn. email: Scientific Reports | (2026) 16:15719 | https://doi.org/10.1038/s41598-026-49611-y 1 interventions. In addition to the specifics of the model, using GPS data has several notable limitations. First, GPS data are highly personalised, which reduces generalisability because models trained on one participant’s data cannot readily be applied to others. Second, GPS collection raises privacy and security concerns due to sensitive location information, meaning many phones operating systems now restrict its collection. Lastly, GPS is power-intensive because it must receive signals from multiple satellites and perform computationally expensive trilateration14, meaning it drains the phone’s battery. By contrast, focusing on high-frequency inertial sensor data (e.g. accelerometers) provide less intrusive, privacy-preserving input that is more generalisable and well-suited for continuous monitoring. Movement data from smartphones and wearable sensors have been previously used by researchers to study and predict smoking behaviour15. Findings suggest that while wearable motion sensors can accurately detect hand-to-mouth gestures in labs, they often perform worse in real-life; possibly due to user behaviour or the need for exact location of device placement (e.g. on the arm). Combining data from smartwatch and a custom wearable finger sensor revealed promising results for detecting smoking-related actions14 (much better than smartwatch alone), however, while relying on wearing multiple devices may be good in lab settings, it is unlikely to work in real-life conditions. Given the increase in smartphones ownership, their improved computing power, and the multiple sensors, such as the accelerometer and GPS, that are now commonly embedded within smartphones, they offer a superior method for passively collecting real-time and objective data. This is particularly attractive if the data from smartphones could be collected in Naturalistic and unconstraint settings. In such settings, the device is not required to be attached to the body, nor is specific training required. Such data have increasingly been used within ML algorithms to analyse and understand complex behavioural patterns to improve insight into different behaviours and medical conditions11. Recent studies have begun to use movement data alone, collected by digital sensors, to predict and understand behaviours in both humans (for example, disruptive behaviour among children (16); calssification of stress or anxietyindicators (17)), and animals (e.g. Ref.18). Furthermore, recent work had shown that some micro-movements, which even when detectable to the human eye often go unnoticed (e (...truncated)