Air pollution is a known risk factor for the exacerbation of many diseases. Among these, is multiple sclerosis (MS), a chronic, autoimmune, neurological disease, characterised by transient episodes of neurological impairment known as relapses. Although the link between environmental factors and relapses has been a subject of investigation in the medical and biostatistical literature, its implications for predictive modelling are still unclear. Thus, in this work, we develop a deep learning model that is able to combine four weeks of environmental data, collected by pollutant-monitoring and weather stations, with patient information to predict an imminent relapse in the following week. Specifically, we cast the task as distinguishing between 4-week sequences followed by a relapse vs. 4-week sequences followed by another relapse-free week, the latter of which were extracted from MS patients who were never observed to have had a relapse. The 1556 sequences were collected in the context of the H2020 BRAINTEASER (”Bringing Artificial Intelligence Home for a Better Care of Amyotrophic Lateral Sclerosis and Multiple Sclerosis”) project. The best-performing model was a recurrent neural network, which yielded an encouraging test-set area under the receiveroperating characteristic curve (AUROC) of 0.70. It also performed adequately (AUROC$=0.60$) on a modified version of the test set where the 4-week relapse-free sequences followed by another relapse-free week were extracted from the same subjects from whom the test sequences followed by a relapse came. Thus, our results, albeit preliminary, suggest that the inclusion of environmental data as the basis of predictive models of MS relapses is a promising direction to obtain short-term predictions, which may be helpful for therapy and life planning. It is especially encouraging that better-than-random performance was preserved on the modified test set, where environmental factors were, by construction, the most informative predictors.

Synthetic clinical data offer several advantages, including the possibility to simulate patient trajectories and investigate long-term outcomes that would otherwise require extensive time and resources to investigate through traditional clinical trials. In this work, we propose a modelling approach based on dynamic Bayesian networks (DBNs) to generate reliable synthetic data that faithfully reproduce the characteristics of original longitudinal clinical datasets. The proposed pipeline includes two main steps: i) learning a DBN model using a layered variable structure informed by domain knowledge, and ii) employing the trained model to simulate longitudinal clinical data. We applied this approach to generate a synthetic version of the LEADER trial dataset, which includes longitudinal data of patients with type 2 diabetes and high cardiovascular risk. After preprocessing, the dataset consisted of 45,556 observations across 82 variables from 8,301 patients. The general utility of the synthetic data was assessed by comparing the distribution of each variable between the synthetic and original datasets. In addition, we evaluated the similarity of Kaplan-Meier survival curves for the major adverse cardiovascular events (MACE), that was the primary endpoints of the trial, between synthetic and real data. Overall, the proposed method demonstrated satisfactory performance, with the synthetic data closely replicating both the variable distributions and time-to-event outcomes observed in the original dataset.

Multiple Sclerosis (MS) is a chronic autoimmune and inflammatory neurological disorder characterised by episodes of symptom exacerbation, known as relapses. Relapses have been linked to environmental factors such as the weather and pollutant concentrations in the air, but the exact relationship between these phenomena is still unclear. In this study, we investigated the role of environmental factors in predicting imminent relapse occurrence in MS patients, leveraging clinical and environmental data collected over a period of one week preceeding the possible event, using data collected in the context of the H2020 BRAINTEASER project. To do this, we developed and tested a range of combinations of predictive models (logistic regression, LR; and random forest, RF) and feature selection schemes, both manual and data-driven. The RF model trained after a data-driven feature selection process based on the Variable Importance in Projection (VIP) metric yielded the best results, i.e., an AUC-ROC of 0.713 and an AUC-PR of 0.639. We identified several key predictors, including clinical variables such as time since MS onset, age at onset, diagnostic delay, and the Expanded Disability Status Scale (EDSS) score, and environmental variables such as wind speed, precipitation, NO2, PM10, average and maximum temperatures, and humidity. These findings suggest that environmental factors may be viable predictors of imminent relapse occurrence in MS.

BACKGROUND: Amyotrophic Lateral Sclerosis (ALS) is a fatal neurodegenerative disorder characterised by the progressive loss of motor neurons in the brain and spinal cord. The fact that ALS's disease course is highly heterogeneous, and its determinants not fully known, combined with ALS's relatively low prevalence, renders the successful application of artificial intelligence (AI) techniques particularly arduous. OBJECTIVE: This systematic review aims at identifying areas of agreement and unanswered questions regarding two notable applications of AI in ALS, namely the automatic, data-driven stratification of patients according to their phenotype, and the prediction of ALS progression. Differently from previous works, this review is focused on the methodological landscape of AI in ALS. METHODS: We conducted a systematic search of the Scopus and PubMed databases, looking for studies on data-driven stratification methods based on unsupervised techniques resulting in (A) automatic group discovery or (B) a transformation of the feature space allowing patient subgroups to be identified; and for studies on internally or externally validated methods for the prediction of ALS progression. We described the selected studies according to the following characteristics, when applicable: variables used, methodology, splitting criteria and number of groups, prediction outcomes, validation schemes, and metrics. RESULTS: Of the starting 1604 unique reports (2837 combined hits between Scopus and PubMed), 239 were selected for thorough screening, leading to the inclusion of 15 studies on patient stratification, 28 on prediction of ALS progression, and 6 on both stratification and prediction. In terms of variables used, most stratification and prediction studies included demographics and features derived from the ALSFRS or ALSFRS-R scores, which were also the main prediction targets. The most represented stratification methods were K-means, and hierarchical and expectation-maximisation clustering; while random forests, logistic regression, the Cox proportional hazard model, and various flavours of deep learning were the most widely used prediction methods. Predictive model validation was, albeit unexpectedly, quite rarely performed in absolute terms (leading to the exclusion of 78 eligible studies), with the overwhelming majority of included studies resorting to internal validation only. CONCLUSION: This systematic review highlighted a general agreement in terms of input variable selection for both stratification and prediction of ALS progression, and in terms of prediction targets. A striking lack of validated models emerged, as well as a general difficulty in reproducing many published studies, mainly due to the absence of the corresponding parameter lists. While deep learning seems promising for prediction applications, its superiority with respect to traditional methods has not been established; there is, instead, ample room for its application in the subfield of patient stratification. Finally, an open question remains on the role of new environmental and behavioural variables collected via novel, real-time sensors.

The organisation of care at diabetes outpatient clinics is typically different from that delivered by general practitioners, it is thus of interest to assess whether there is also a difference in the predictive power of heart failure hospitalisation risk scores developed independently for each subpopulation. To such a purpose, a diabetes outpatient clinic dataset and a primary care dataset were considered. A Cox proportional hazard model, an accelerated failure time model, a logistic regression, a random forest, and a K-nearest neighbours model were trained in each dataset and tested on both. The UK Prospective Diabetes Study (UKPDS) risk engine was used as benchmark. Results show that models developed using primary care data performed well on the corresponding test set but poorly when used in the diabetes outpatient clinic setting (best C-Index = 0.759vs. 0.615, best AUROC = 0.757vs. 0.598). Models trained on the diabetes outpatient clinic data performed well on the corresponding test set, and their predictive power in the primary care setting was not statistically different from the one of models developed using primary care data (best C-Index = 0.814 vs 0.740, best AUROC = 0.812 vs 0.750). In both settings UKPDS had lower predictive power than the best newly-developed models. Different care setting led to a difference in the predictive power of heart failure hospitalisation risk scores that depended on both the data used for training and the methodological approach chosen. This suggests the need to consider these factors when applying risk scores to a target population where the expected incidence of the outcome and the distribution of baseline covariates differ from those of the population for which scores were proposed.

People with diabetes require lifelong access to healthcare services to delay the onset of complications. Their disease management processes generate great volumes of data across several domains, from clinical to administrative. Difficulties in accessing and processing these data hinder their secondary use in an institutional setting, even for highly desirable applications, such as the prediction of cardiovascular disease, the main driver of excess mortality in diabetes. Hence, in the present work, we propose a deep learning model for the prediction of major adverse cardiovascular events (MACE), developed and validated using the administrative claims of 214,676 diabetic patients of the Veneto region, in North East Italy. Specifically, we use a year of pharmacy and hospitalisation claims, together with basic patient's information, to predict the 4P-MACE composite endpoint, i.e., the first occurrence of death, heart failure, myocardial infarction, or stroke, with a variable prediction horizon of 1 to 5 years. Adapting to the time-to-event nature of this task, we cast our problem as a multi-outcome (4P-MACE and components), multi-label (1 to 5 years) classification task with a custom loss to account for the effect of censoring. Our model, purposefully specified to minimise data preparation costs, exhibits satisfactory performance in predicting 4P-MACE at all prediction horizons: AUROC from 0.812 (C.I.: 0.797 - 0.827) to 0.792 (C.I.: 0.781 - 0.802); C-index from 0.802 (C.I.: 0.788 - 0.816) to 0.770 (C.I.: 0.761 - 0.779). Components' prediction performance is also adequate, ranging from death's 0.877 1-year AUROC to stroke's 0.689 5-year AUROC.