Publications
2025
Rahmandad, Hazhir; Akhavan, Ali; Jalali, Mohammad S
Incorporating Deep Learning Into System Dynamics: Amortized Bayesian Inference for Scalable Likelihood‐Free Parameter Estimation Journal Article
In: System Dynamics Review, vol. 41, iss. 1, 2025.
Abstract | Links | BibTeX | Tags: Artificial intelligence, Methods
@article{Rahmandad2025,
title = {Incorporating Deep Learning Into System Dynamics: Amortized Bayesian Inference for Scalable Likelihood‐Free Parameter Estimation},
author = {Hazhir Rahmandad and Ali Akhavan and Mohammad S Jalali
},
url = {https://mj-lab.mgh.harvard.edu/wp-content/uploads/2025/01/Rahmandad_Amortized-Bayesian_2025.pdf},
year = {2025},
date = {2025-01-24},
journal = {System Dynamics Review},
volume = {41},
issue = {1},
abstract = {Estimating parameters and their credible intervals for complex system dynamics models is challenging but critical to continuous model improvement and reliable communication with an increasing fraction of audiences. The purpose of this study is to integrate Amortized Bayesian Inference (ABI) methods with system dynamics. Utilizing Neural Posterior Estimation (NPE), we train neural networks using synthetic data (pairs of ground truth parameters and outcome time series) to estimate parameters of system dynamics models. We apply this method to two example models: a simple Random Walk model and a moderately complex SEIRb model. We show that the trained neural networks can output the posterior for parameters instantly given new unseen time series data. Our analysis highlights the potential of ABI to facilitate a principled, scalable, and likelihood-free inference workflow that enhance the integration of models of complex systems with data. Accompanying code streamlines application to diverse system dynamics models.},
keywords = {Artificial intelligence, Methods},
pubstate = {published},
tppubtype = {article}
}
Estimating parameters and their credible intervals for complex system dynamics models is challenging but critical to continuous model improvement and reliable communication with an increasing fraction of audiences. The purpose of this study is to integrate Amortized Bayesian Inference (ABI) methods with system dynamics. Utilizing Neural Posterior Estimation (NPE), we train neural networks using synthetic data (pairs of ground truth parameters and outcome time series) to estimate parameters of system dynamics models. We apply this method to two example models: a simple Random Walk model and a moderately complex SEIRb model. We show that the trained neural networks can output the posterior for parameters instantly given new unseen time series data. Our analysis highlights the potential of ABI to facilitate a principled, scalable, and likelihood-free inference workflow that enhance the integration of models of complex systems with data. Accompanying code streamlines application to diverse system dynamics models.
2024
Mahmoudi, Hesam; Chang, Doris; Lee, Hannah; Ghaffarzadegan, Navid; Jalali, Mohammad S.
A Critical Assessment of Large Language Models for Systematic Reviews: Utilizing ChatGPT for Complex Data Extraction Working paper
2024.
BibTeX | Tags: Artificial intelligence
@workingpaper{Mahmoudi2024,
title = {A Critical Assessment of Large Language Models for Systematic Reviews: Utilizing ChatGPT for Complex Data Extraction},
author = {Hesam Mahmoudi and Doris Chang and Hannah Lee and Navid Ghaffarzadegan and Mohammad S. Jalali},
year = {2024},
date = {2024-04-15},
keywords = {Artificial intelligence},
pubstate = {published},
tppubtype = {workingpaper}
}
Akhavan, Ali; Jalali, Mohammad S.
Generative AI and Simulation Modeling: How Should You (Not) Use Large Language Models Like ChatGPT Journal Article
In: System Dynamics Review, 2024.
Abstract | Links | BibTeX | Tags: Artificial intelligence, Simulation modeling
@article{nokey,
title = {Generative AI and Simulation Modeling: How Should You (Not) Use Large Language Models Like ChatGPT },
author = {Ali Akhavan and Mohammad S. Jalali },
url = {https://mj-lab.mgh.harvard.edu/wp-content/uploads/2024/07/Generative_AI_sim_mod.pdf},
year = {2024},
date = {2024-01-18},
urldate = {2024-01-18},
journal = {System Dynamics Review},
abstract = {Generative Artificial Intelligence (AI) tools, such as Large Language Models (LLMs) and chatbots like ChatGPT, hold promise for advancing simulation modeling in various domains. Despite their growing prominence and associated debates, there remains a gap in comprehending the potential of generative AI in this field and a lack of guidelines for its effective deployment. This paper endeavors to bridge these gaps. We discuss the applications of ChatGPT through an example of modeling COVID-19’s impact on economic growth in the United States. Although we utilize ChatGPT, our guidelines are generic and can be applied to a broader range of generative AI tools and platforms. Our work presents a systematic approach for integrating generative AI across the simulation research continuum, from problem articulation to insight derivation and documentation, independent of the specific simulation modeling method chosen. We emphasize that while these tools offer enhancements in refining ideas and expediting processes, they should complement rather than replace critical thinking inherent to research.},
keywords = {Artificial intelligence, Simulation modeling},
pubstate = {published},
tppubtype = {article}
}
Generative Artificial Intelligence (AI) tools, such as Large Language Models (LLMs) and chatbots like ChatGPT, hold promise for advancing simulation modeling in various domains. Despite their growing prominence and associated debates, there remains a gap in comprehending the potential of generative AI in this field and a lack of guidelines for its effective deployment. This paper endeavors to bridge these gaps. We discuss the applications of ChatGPT through an example of modeling COVID-19’s impact on economic growth in the United States. Although we utilize ChatGPT, our guidelines are generic and can be applied to a broader range of generative AI tools and platforms. Our work presents a systematic approach for integrating generative AI across the simulation research continuum, from problem articulation to insight derivation and documentation, independent of the specific simulation modeling method chosen. We emphasize that while these tools offer enhancements in refining ideas and expediting processes, they should complement rather than replace critical thinking inherent to research.
Jalali, Mohammad S.; Akhavan, Ali
Integrating AI Language Models in Qualitative Research: Replicating Interview Data Analysis with ChatGPT Journal Article
In: System Dynamics Review, 2024.
Abstract | Links | BibTeX | Tags: Artificial intelligence
@article{nokey,
title = {Integrating AI Language Models in Qualitative Research: Replicating Interview Data Analysis with ChatGPT},
author = {Mohammad S. Jalali and Ali Akhavan},
url = {https://mj-lab.mgh.harvard.edu/wp-content/uploads/2024/07/Integrating_AI_language_models_qual_res.pdf},
year = {2024},
date = {2024-01-17},
urldate = {2024-01-17},
journal = {System Dynamics Review},
abstract = {The recent advent of artificial intelligence (AI) language tools like ChatGPT has opened up new opportunities in qualitative research. We revisited a previous project on obesity prevention interventions, where we developed a causal loop diagram through in-depth interview data analysis. Utilizing ChatGPT in our replication process, we compared its results against our original approach. We discuss that ChatGPT contributes to improved efficiency and unbiased data processing; however, it also reveals limitations in context understanding. Our findings suggest that AI language tools currently have great potential to serve as an augmentative tool rather than a replacement for the intricate analytical tasks performed by humans. With ongoing advancements, AI technologies may soon offer more substantial support in enhancing qualitative research capabilities, an area that deserves more investigation. },
keywords = {Artificial intelligence},
pubstate = {published},
tppubtype = {article}
}
The recent advent of artificial intelligence (AI) language tools like ChatGPT has opened up new opportunities in qualitative research. We revisited a previous project on obesity prevention interventions, where we developed a causal loop diagram through in-depth interview data analysis. Utilizing ChatGPT in our replication process, we compared its results against our original approach. We discuss that ChatGPT contributes to improved efficiency and unbiased data processing; however, it also reveals limitations in context understanding. Our findings suggest that AI language tools currently have great potential to serve as an augmentative tool rather than a replacement for the intricate analytical tasks performed by humans. With ongoing advancements, AI technologies may soon offer more substantial support in enhancing qualitative research capabilities, an area that deserves more investigation.
2023
Tatar, Moosa; Faraji, Mohammad R.; Keyes, Katherine; Wilson, Fernando A.; Jalali, Mohammad S.
Social Vulnerability Predictors of Drug Poisoning Mortality: A Machine Learning Analysis in the United States Journal Article
In: The American Journal on Addictions, 2023.
Abstract | Links | BibTeX | Tags: Artificial intelligence, Disparity and equity, Substance use
@article{669757,
title = {Social Vulnerability Predictors of Drug Poisoning Mortality: A Machine Learning Analysis in the United States},
author = {Moosa Tatar and Mohammad R. Faraji and Katherine Keyes and Fernando A. Wilson and Mohammad S. Jalali},
url = {https://mj-lab.mgh.harvard.edu/wp-content/uploads/2023/07/Tatar_et_al_AJA_2023.pdf},
year = {2023},
date = {2023-05-01},
urldate = {2023-05-01},
journal = {The American Journal on Addictions},
abstract = {Background and Objectives
Drug poisoning is a leading cause of unintentional deaths in the United States. Despite the growing literature, there are a few recent analyses of a wide range of community-level social vulnerability features contributing to drug poisoning mortality. Current studies on this topic face three limitations: often studying a limited subset of vulnerability features, focusing on small sample sizes, or solely including local data. To address this gap, we conducted a national-level analysis to study the impacts of several social vulnerability features in predicting drug mortality rates in the United States.
Methods
We used machine learning to investigate the role of 16 social vulnerability features in predicting drug mortality rates for US counties in 2014, 2016, and 2018—the most recent available data. We estimated each vulnerability feature's gain relative contribution in predicting drug poisoning mortality.
Results
Among all social vulnerability features, the percentage of noninstitutionalized persons with a disability is the most influential predictor, with a gain relative contribution of 18.6%, followed by population density and the percentage of minority residents (13.3% and 13%, respectively). Percentages of households with no available vehicles, mobile homes, and persons without a high school diploma are the following features with gain relative contributions of 6.3%, 5.8%, and 5.1%, respectively.
Conclusion and Scientific Significance
We identified social vulnerability features that are most predictive of drug poisoning mortality. Public health interventions and policies targeting vulnerable communities may increase the resilience of these communities and mitigate the overdose death and drug misuse crisis.},
keywords = {Artificial intelligence, Disparity and equity, Substance use},
pubstate = {published},
tppubtype = {article}
}
Background and Objectives
Drug poisoning is a leading cause of unintentional deaths in the United States. Despite the growing literature, there are a few recent analyses of a wide range of community-level social vulnerability features contributing to drug poisoning mortality. Current studies on this topic face three limitations: often studying a limited subset of vulnerability features, focusing on small sample sizes, or solely including local data. To address this gap, we conducted a national-level analysis to study the impacts of several social vulnerability features in predicting drug mortality rates in the United States.
Methods
We used machine learning to investigate the role of 16 social vulnerability features in predicting drug mortality rates for US counties in 2014, 2016, and 2018—the most recent available data. We estimated each vulnerability feature's gain relative contribution in predicting drug poisoning mortality.
Results
Among all social vulnerability features, the percentage of noninstitutionalized persons with a disability is the most influential predictor, with a gain relative contribution of 18.6%, followed by population density and the percentage of minority residents (13.3% and 13%, respectively). Percentages of households with no available vehicles, mobile homes, and persons without a high school diploma are the following features with gain relative contributions of 6.3%, 5.8%, and 5.1%, respectively.
Conclusion and Scientific Significance
We identified social vulnerability features that are most predictive of drug poisoning mortality. Public health interventions and policies targeting vulnerable communities may increase the resilience of these communities and mitigate the overdose death and drug misuse crisis.
Drug poisoning is a leading cause of unintentional deaths in the United States. Despite the growing literature, there are a few recent analyses of a wide range of community-level social vulnerability features contributing to drug poisoning mortality. Current studies on this topic face three limitations: often studying a limited subset of vulnerability features, focusing on small sample sizes, or solely including local data. To address this gap, we conducted a national-level analysis to study the impacts of several social vulnerability features in predicting drug mortality rates in the United States.
Methods
We used machine learning to investigate the role of 16 social vulnerability features in predicting drug mortality rates for US counties in 2014, 2016, and 2018—the most recent available data. We estimated each vulnerability feature's gain relative contribution in predicting drug poisoning mortality.
Results
Among all social vulnerability features, the percentage of noninstitutionalized persons with a disability is the most influential predictor, with a gain relative contribution of 18.6%, followed by population density and the percentage of minority residents (13.3% and 13%, respectively). Percentages of households with no available vehicles, mobile homes, and persons without a high school diploma are the following features with gain relative contributions of 6.3%, 5.8%, and 5.1%, respectively.
Conclusion and Scientific Significance
We identified social vulnerability features that are most predictive of drug poisoning mortality. Public health interventions and policies targeting vulnerable communities may increase the resilience of these communities and mitigate the overdose death and drug misuse crisis.
2022
Stafford, Celia; Marrero, Wesley; Naumann, Rebecca B; Lich, Kristen Hassmiller; Wakeman, Sarah; Jalali, Mohammad S.
Identifying Key Risk Factors for Premature Discontinuation of Opioid Use Disorder Treatment in the United States: a Predictive Modeling Study Journal Article
In: Drug and Alcohol Dependence, vol. 237, no. 1, pp. 109507, 2022.
Abstract | Links | BibTeX | Tags: Artificial intelligence, Substance use
@article{669758,
title = {Identifying Key Risk Factors for Premature Discontinuation of Opioid Use Disorder Treatment in the United States: a Predictive Modeling Study},
author = {Celia Stafford and Wesley Marrero and Rebecca B Naumann and Kristen Hassmiller Lich and Sarah Wakeman and Mohammad S. Jalali},
url = {https://mj-lab.mgh.harvard.edu/wp-content/uploads/2022/12/Stafford_2022_DAD.pdf},
doi = {10.1016/j.drugalcdep.2022.109507},
year = {2022},
date = {2022-06-01},
urldate = {2022-06-01},
journal = {Drug and Alcohol Dependence},
volume = {237},
number = {1},
pages = {109507},
abstract = {Background
Treatment for opioid use disorder (OUD), particularly medication for OUD, is highly effective; however, retention in OUD treatment is a significant challenge. We aimed to identify key risk factors for premature exit from OUD treatment.
Methods
We analyzed 2,381,902 cross-sectional treatment episodes for individuals in the U.S., discharged between Jan/1/2015 and Dec/31/2019. We developed classification models (Random Forest, Classification and Regression Trees (CART), Bagged CART, and Boosted CART), and analyzed 31 potential risk factors for premature treatment exit, including treatment characteristics, substance use history, socioeconomic status, and demographic characteristics. We stratified our analysis based on length of stay in treatment and service setting. Models were compared using cross-validation and the receiver operating characteristic area under the curve (ROC-AUC).
Results
Random Forest outperformed other methods (ROC-AUC: 74%). The most influential risk factors included characteristics of service setting, geographic region, primary source of payment, and referral source. Race, ethnicity, and sex had far weaker predictive impacts. When stratified by treatment setting and length of stay, employment status and delay (days waited) to enter treatment were among the most influential factors. Their importance increased as treatment duration decreased. Notably, importance of referral source increased as the treatment duration increased. Finally, age and age of first use were important factors for lengths of stay of 2–7 days and in detox treatment settings.
Conclusions
The key factors of OUD treatment attrition identified in this analysis should be more closely explored (e.g., in causal studies) to inform targeted policies and interventions to improve models of care.},
keywords = {Artificial intelligence, Substance use},
pubstate = {published},
tppubtype = {article}
}
Background
Treatment for opioid use disorder (OUD), particularly medication for OUD, is highly effective; however, retention in OUD treatment is a significant challenge. We aimed to identify key risk factors for premature exit from OUD treatment.
Methods
We analyzed 2,381,902 cross-sectional treatment episodes for individuals in the U.S., discharged between Jan/1/2015 and Dec/31/2019. We developed classification models (Random Forest, Classification and Regression Trees (CART), Bagged CART, and Boosted CART), and analyzed 31 potential risk factors for premature treatment exit, including treatment characteristics, substance use history, socioeconomic status, and demographic characteristics. We stratified our analysis based on length of stay in treatment and service setting. Models were compared using cross-validation and the receiver operating characteristic area under the curve (ROC-AUC).
Results
Random Forest outperformed other methods (ROC-AUC: 74%). The most influential risk factors included characteristics of service setting, geographic region, primary source of payment, and referral source. Race, ethnicity, and sex had far weaker predictive impacts. When stratified by treatment setting and length of stay, employment status and delay (days waited) to enter treatment were among the most influential factors. Their importance increased as treatment duration decreased. Notably, importance of referral source increased as the treatment duration increased. Finally, age and age of first use were important factors for lengths of stay of 2–7 days and in detox treatment settings.
Conclusions
The key factors of OUD treatment attrition identified in this analysis should be more closely explored (e.g., in causal studies) to inform targeted policies and interventions to improve models of care.
Treatment for opioid use disorder (OUD), particularly medication for OUD, is highly effective; however, retention in OUD treatment is a significant challenge. We aimed to identify key risk factors for premature exit from OUD treatment.
Methods
We analyzed 2,381,902 cross-sectional treatment episodes for individuals in the U.S., discharged between Jan/1/2015 and Dec/31/2019. We developed classification models (Random Forest, Classification and Regression Trees (CART), Bagged CART, and Boosted CART), and analyzed 31 potential risk factors for premature treatment exit, including treatment characteristics, substance use history, socioeconomic status, and demographic characteristics. We stratified our analysis based on length of stay in treatment and service setting. Models were compared using cross-validation and the receiver operating characteristic area under the curve (ROC-AUC).
Results
Random Forest outperformed other methods (ROC-AUC: 74%). The most influential risk factors included characteristics of service setting, geographic region, primary source of payment, and referral source. Race, ethnicity, and sex had far weaker predictive impacts. When stratified by treatment setting and length of stay, employment status and delay (days waited) to enter treatment were among the most influential factors. Their importance increased as treatment duration decreased. Notably, importance of referral source increased as the treatment duration increased. Finally, age and age of first use were important factors for lengths of stay of 2–7 days and in detox treatment settings.
Conclusions
The key factors of OUD treatment attrition identified in this analysis should be more closely explored (e.g., in causal studies) to inform targeted policies and interventions to improve models of care.