I) Artificial Intelligence/ Machine Learning:
Artificial Neural Networks: Theory and Applications
The use of artificial neural networks in medical field has grown rapidly in recent years. Neural networks have been combined with traditional statistical models, including linear, logistic, and multinomial logistic regression. However, important challenges remain, particularly when outcomes do not follow normal or binomial distributions, such as count data. This motivated me to develop approaches that combine the flexibility of neural networks with the statistical foundation of traditional regression models.
We developed a series of hybrid neural-network extensions of generalized linear models (GLM) that replace the traditional linear predictor with a feed-forward neural network. For count outcomes, this led to Poisson regression Neural Networks (PNN) and, more recently, Negative Binomial Neural Networks (NBNN). We then extended the same framework to continuous positive outcomes using a gamma distribution, resulting in Gamma Neural Networks (GNN). These models can capture non-linear effects and interactions directly from the data without requiring researchers to specify complex higher-order terms in advance, while maintaining the familiar probability distributions and likelihood-based framework of traditional regression models.
The theoretical development of this framework has been published, and we have demonstrated its application in cardiovascular disease and Alzheimer’s disease research. More recently, we finalized the NBNN and GNN models, and the corresponding manuscript is published. Overall, this work aims to bridge traditional statistical modeling and modern neural-network methods by providing flexible approaches for analyzing complex medical data while retaining a strong statistical foundation and meaningful interpretation.
Non-linear Poisson Regression Using Neural Networks: A Simulation Study. Neural Computing and Applications 2009;18(8):939-49.
Nonlinear Association between Serum Testosterone Levels and Coronary Artery Disease in Iranian Men. European Journal of Epidemiology 2009;24(6):297-306 (application of the neural Poisson model).
Applying Neural Network Poisson Regression to Predict Cognitive Score Changes. Journal of Applied Statistics 2011;38(9):2051-62.
A Machine Learning Approach to Modify the Neurocognitive Frailty Index for the Prediction of Cognitive Status in the Canadian Population. Journal of Clinical Medicine 2025; 14 (18), 6509
Developing Nonlinear Negative Binomial and Non-linear Gamma Regressions and Non-linear Poisson Regression Using Neural Networks: A Simulation Study. Neural Computing and Applications 2026
Markov Models in Cognitive Psychology and Neuroscience Studies
In my postdoctoral training I developed and applied another machine learning method (multi-state transition model) for cognitive functions assessment and healthy aging. The model was based on the Markov process with interpretable parameters which could be adjusted for covariates. The model can compute probabilities of improvements, stabilization, worsening and dying as a function of baseline and the covariates estimate according to modified Poisson model between two assessments. This novel multi-state transition model has brought several opportunities for collaboration and publication, so far model was used and published in the top epidemiology and statistical journals such as Neuroepidemiology 2009, Annals of Epidemiology (3 times) 2010, 2011, 2013; and Statistical Methods in Medical Research, 2014.
See below link for more technical explanations of multi-state transition model:
A brief explanation about Multi-state transition model
Application of Machine Learning in Health and Medicine
The use of machine learning techniques has received much attention for the purposes of analyzing large medical database that may contain several hundred variables and many thousands (or perhaps millions) of participants. These conditions are addressed
by machine learning algorithms that identify key variables needed in a predictive model. In addition, a variety of techniques are used to locate nonlinear relationships and interactive effects. I applied machine learning and Artificial Intelligence methods in neurological disease (Alzheimer’s disease & spinal cord injury) to accommodate the non-linear nature of data. Specifically, I utilize machine learning methodologies to develop clinical tools aimed at informing patient outcomes such as neurological and functional recovery, care and the quality of life of individuals living with spinal cord injury. I hope to develop interactive and interpretable artificial intelligence algorithms that are potentially applicable to several neurological health conditions, in addition to SCI. Some recent works are listed below:
A simplified clinical prediction rule for prognosticating independent walking after spinal cord injury: a prospective study from a Canadian multicenter spinal cord injury registry. (The Spine Journal, 2017)
Predicting Injury Severity and Neurological Recovery after Acute Cervical Spinal Cord Injury: A Comparison of Cerebrospinal Fluid and Magnetic Resonance Imaging Biomarkers. (Journal of Neurotrauma, 2018)
Unbiased Recursive Partitioning to Stratify Patients with Acute Traumatic Spinal Cord Injuries: External Validity in an Observational Cohort Study. (Journal of Neurotrauma, 2019)
Highlighting discrepancies in walking prediction accuracy for patients with traumatic spinal cord injury: an evaluation of validated prediction models using a Canadian Multicenter Spinal Cord Injury Registry. (The Spine Journal, 2019)
Decision tree analysis to better control treatment effects in spinal cord injury clinical research. (The Journal of Neurosurgery, 2019)
Development of an unsupervised machine learning algorithm for the prognostication of walking ability in spinal cord injury patients ( The Spine Journal, 2020)
Development of a Machine Learning Algorithm for Predicting In-hospital and One-year Mortality after Traumatic Spinal Cord Injury: Mortality prediction tool for Spinal Cord Injury (The Spine Journal, 2022)
II) Aging, Frailty, Alzheimer’s disease and Dementia:
Chronological Age vs. Biological Age
Chronological age is very important factor of aging process; however, people with same age are dramatically different in activity of daily living, quality of life and survival. So many efforts were done to understand association between biological age and frailty especially high-risk people for dementia. Frailty is an age-related vulnerability state created by a multidimensional loss of reserves (energy, physical ability, cognition and general health). Elderly people who are frail are more prone to adverse outcomes than elderly people who are fit (less deficit count). There are many scales to measure frailty; Mitnitski and Rockwood (Scientific World Journal. 2001) have developed a framework based on number of deficit.
I was involved to and applying frailty index in several projects using a well-known Canadian population survey (CSHA: Canadian Study of Health and Aging). Moreover I also involved, applying frailty index in an American prospective cohort study of elderly people living in King County (Seattle), WA study (Annals of Epidemiology. 2010) and in the Yale Precipitating Events Project, New Haven Connecticut (Journal of American Geriatric Society. 2011). This index was also validated in two European Studies. The first study being an analysis of the cohort of the Oxford Project To Investigate Memory and Aging (OPTIMA), a longitudinal observational cohort (Age Ageing. 2013) and the second was a study on the elderly rural Italy (J Nutr Health Aging. 2010). I aim to continue research on this area by developing other type of frailty index for a specific health condition, such as spinal cord injury.
Developing a Screening Tool for Dementia
It is estimated almost 750,000 Canadians are living with cognitive impairment or dementia. Among older adults diagnosed with mild cognitive impairment, approximately 30 percent will develop dementia within five years. However still we don’t have good screening tool to detect Alzheimer’s and dementia at early stage.
It is known that the frailty index can remarkably predict the physical health of people; however, one of our previous research studies we have shown that there is a room for adding cognitive domain (Age Ageing. 2013) to boost this tool for a broader purpose (Toward making a globally acceptable screening tool). The majority of studies to date have used the frailty index as a measure of physical fitness but mental fitness (mental health) receive less attention in studies of older adults. As a result, there is little understanding of the variability in physical and cognitive change in relation to Alzheimer’s disease and mortality.
Recently, a novel Neurocognitive Frailty Index (NFI) was developed by adding several cognitive domains to the Frailty Index. The addition of cognitive components to frailty (index/scale) made it more comprehensive measure to assess frailty in the elderly. NFI as a screening tool provides a helpful measure to assess frailty, in fact researchers merged both physical and mental components to construct the index. The result suggests these symptoms should be considered to better encompass the complexity of the frailty syndrome. This research can apply the knowledge about the stochastic nature of cognitive function, its relation with any type of dementia specifically AD and significantly advance our knowledge on the exact impact of cognitive change on brain health and survival in high risk people.
III) Pharmacoepidemiology and Therapeutic Initiative:
Pharmacoepidemiology is the study of the utilization and effects of drugs/devices in population, it provides an estimate of the probability of beneficial effects of a drug in a population and the probability of adverse effects. Pharmacoepidemiology concentrates on patient outcomes from therapeutics by using methods of epidemiology, biostatistics and computer sciences.
One challenge in pharmacoepidemiology is how to avoid the biases that may arise in estimating the strength of association between a drug and an adverse event, because the data available are not usually obtained under experimental conditions. In this situation, it may be difficult to select suitable comparison groups among those who are not taking the drug. In pharmacoepidemiology, existing statistical methods can handle confounders if they are recognized and measured in an accurate way. However most of the time, the confounds are unknown, unmeasured, or imperfectly measured, or the relationship between variables does not follow a simple linear pattern.
There have been numerous clinical trials conducted over past 30 years, but little progress has been made in identifying clinically effective spinal cord injury treatments. The question arises whether successful therapies have been eluded us simply because these potential treatments haven’t worked, or if the methodological approach taken by these trials is imprecise. It is necessary to identify and quantify variables that strongly influence outcome in order to control for bias. It will increase the likelihood that a statistically significant effect belongs to a real effect by improve the power of studies, and most importantly, increase the likelihood of detecting a true therapeutic effect.
We have shown in a series of publications that by optimizing study design and analyzing clinical data comprehensively, it is possible to evaluate the impact of interventions like surgery (e.g. time of surgery), rehabilitation (e.g. intensity), medical device (e.g. new vs. used catheter), and medication (e.g. therapeutic evaluation of methylprednisolone).