AIMI: Leveraging Future Knowledge and Personalization in Sparse Event Forecasting for Treatment Adherence

Abdullah Mamun, Diane J. Cook, Hassan Ghasemzadeh · 2025

Read AIMI: Leveraging Future Knowledge and Personalization in Sparse Event Forecasting for Treatment Adherence by Abdullah Mamun, Diane J. Cook, Hassan Ghasemzadeh (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-16091.txt.

Introduction

Abdullah Mamun, Diane Cook, and Hassan Ghasemzadeh built a forecaster for missed medication in a 2025 paper. People with chronic disease who skip doses face costly harm, and timely nudges could help if risk were known in advance. Phones and wearables now make such forecasting practical.

Their system, AIMI, combines phone sensor data with each person's medication history to estimate the chance of forgetting a dose. In a study with 27 heart patients, their best network reached about 93 percent accuracy. Tests stripping parts away showed that two ingredients matter: using what is known about the future schedule, and training tuned to the person. The authors release the code for others to build on.

Catalog record

CREATORAbdullah Mamun, Diane J. Cook, Hassan Ghasemzadeh
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-16091

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AIMI: Leveraging Future Knowledge and Personalization in Sparse Event Forecasting for Treatment Adherence Abdullah Mamun, Diane J. Cook, Hassan Ghasemzadeh Abstract Adherence to prescribed treatments is crucial for individuals with chronic conditions to avoid costly or adverse health outcomes. For certain patient groups, intensive lifestyle interventions are vital for enhancing medication adherence. Accurate forecasting of treatment adherence can open pathways to developing an on-demand intervention tool, enabling timely and personalized support. With the increasing popularity of smartphones and wearables, it is now easier than ever to develop and deploy smart activity monitoring systems. However, effective forecasting systems for treatment adherence based on wearable sensors are still not widely available. We close this gap by proposing A dherence Forecasting and I ntervention with M achine I ntelligence (AIMI). AIMI is a knowledge-guided adherence forecasting system that leverages smartphone sensors and previous medication history to estimate the likelihood of forgetting to take a prescribed medication. A user study was conducted with 27 participants who took daily medications to manage their cardiovascular diseases. We designed and developed CNN and LSTM-based forecasting models with various combinations of input features and found that LSTM models can forecast medication adherence with an accuracy of 0.932 and an F-1 score of 0.936. Moreover, through a series of ablation studies involving convolutional and recurrent neural network architectures, we demonstrate that leveraging known knowledge about future and personalized training enhances the accuracy of medication adherence forecasting. Code available: https://github.com/ab9mamun/AIMI. 1 Introduction Approximately 6.7 million adults are affected by heart failure (HF) in the U.S. [ 40 ] . The number of American adults with HF has increased by 31.4% in the last 10 years [ 17 ] . It is a deadly disease with a 5-year mortality rate of 50% and a 10-year excess mortality rate of 36-40% [ 52 , 14 ] . Medication adherence is crucial for many HF patients as well as individuals suffering from hypertension, hyperlipidemia, and diabetes [ 47 , 20 , 15 , 5 ] . Despite the risks, people often fail to adhere to the prescribed medications [ 45 , 27 ] . Lifestyle interventions, such as smartphone apps and reminders, can improve medication adherence, but low-cost reminder solutions are not as effective as intensive interventions in increasing medication adherence [ 3 , 26 , 11 , 26 ] . In this study, we aim to design an effective intervention system by accurately forecasting whether a person is likely to miss their prescribed medication. Treatment adherence forecasting was addressed in different prior publications in different application domains. An adherence prediction system with Long Short-Term Memory (LSTM) networks was developed in [ 18 ] that can predict adherence to specific injection-based treatments with 77.35% accuracy. In that study, Internet-connected sharps bins were used to collect treatment events. Another study developed adherence forecasting among individuals with psychiatric disorders with an AUC of 0.87 [ 28 ] . Several publications from the past addressed the forecasting of health events such as hospital readmission, development of heart failure, and other adverse health outcomes [ 10 , 41 , 55 ] . Disease progression and prediction of health outcomes were previously addressed using electronic health records (EHR) [ 1 ] . LSTM-based models were proposed for forecasting daily step counts [ 36 , 37 ] . [ 21 ] forecasts the medical service demand with a hybrid model based on ARIMA and a self-adaptive filtering method. Using future knowledge to improve prediction is a relatively new idea, and prior work in this area is very limited [ 46 ] . Healthcare datasets collected in the wild face difficulties related to data noise, missing data, and natural human variance [ 39 ] . In this case, any additional available information should be considered for inclusion in the feature set. However, current literature does not provide enough evidence to draw a reliable conclusion about the impact of factors such as time context, location, or future knowledge—i.e., known future events—on a person’s treatment adherence. We hypothesize that medication adherence can be affected by a person’s temporal and behavioral context, such as the day of the week, location, and activity. Consider the situation where someone is in the middle of a trip or went to bed later than usual, possibly because they are sleeping late the next morning. In these two cases, they may not take the medication at the prescribed time. Although adherence forecasting using past electronic health records has been explored in the past [ 29 ] , forecasting such events using sensor or physical activity data or future knowledge features did not get sufficient attention. This research aims to bridge this gap by formulating a sparse event forecasting problem and exploring context-aware and knowledge-guided deep learning solutions. We propose a smart health system, AIMI( A dherence Forecasting and I ntervention with M achine I ntelligence), that considers not only the current activity and history of adherence but also the time-context and known information of future prescriptions. To evaluate our method, we conducted a user study with N=27 participants who take medications to manage cardiovascular diseases. Then Convolutional Neural Networks (CNN) and LSTM models were trained and evaluated as potential forecasting models for the AIMI system. Furthermore, we include an incremental learning algorithm for training the system in a resource-constrained training environment, e.g., limited memory. This incremental learning not only facilitates on-device learning in healthcare systems but also ensures the security of the data as it eliminates the need to transfer any data to third-party servers to train neural networks. In such training settings, the performance of neural networks is prone to worsen for specific instances after they are trained on new instances [ 9 ] which we overcome with personalization of the final model. In summary, our main contributions are: i) investigating the challenges of predicting and forecasting treatment adherence, ii) proposing the AIMI system for forecasting treatment adherence using sensor data, adherence history, and future knowledge features, iii) providing implementations of our proposed method using CNN and LSTM models and a systematic comparison of the models for this task, iv) conducting a user study and evaluating the AIMI on the collected dataset, v) proposing an incremental learning algorithm for resource-constrained environments, and vi) determining the impact of knowledge guidance and personalization of machine learning models for treatment adherence forecasting. 2 Background and Related Work 2.1 Treatment Adherence and Personalization Treatment adherence can be defined in a few possible ways. One form of nonadherence is not taking the medication at the prescribed time. Whether a person will adhere to the prescribed medications can be predicted with baseline questionnaires [ 44 ] . Another form of nonadherence can be violating a medical temporal constraint (MTC) associated with the medicine, such as a recommendation to take a particular medication after a meal or before a meal. [ 50 ] predicts the violation of a medical temporal constraint with a behavioral prediction model, HERBERT, that utilizes a personalized training process. Predicting or forecasting the health events of a person differs from the same statistical variables that are affected by the actions of multiple entities of the world. A generic classifier may achieve strong performance on one set of participants’ data but

Read the complete text