BACKGROUND: Current clinical decision support systems (DSS) are trained and validated on observational data from the clinic in which the DSS is going to be applied. This is problematic for treatments that have already been validated in a randomized clinical trial (RCT), but have not yet been introduced in any clinic. In this work, we report on a method for training and validating the DSS core before introduction to a clinic, using the RCT data themselves. The key challenges we address are of missingness, foremost: missing rationale when assigning a treatment to a patient (the assignment is at random), and missing verification evidence, since the effectiveness of a treatment for a patient can only be verified (ground truth) if the treatment was indeed assigned to the patient - but then the assignment was at random. MATERIALS: We use the data of a multi-armed clinical trial that investigated the effectiveness of single treatments and combination treatments for 240+ tinnitus patients recruited and treated in 5 clinical centres. METHODS: To deal with the 'missing rationale for treatment assignment' challenge, we re-model the target variable that measures the outcome of interest, in order to suppress the effect of the individual treatment, which was at random, and control on the effect of treatment in general. To deal with missing features for many patients, we use a learning core that is robust to missing features. Further, we build ensembles that parsimoniously exploit the small patient numbers we have for learning. To deal with the 'missing verification evidence' challenge, we introduce counterfactual treatment verification, a verification scheme that juxtaposes the effectiveness of the recommendations of our approach to the effectiveness of the RCT assignments in the cases of agreement/disagreement between the two. RESULTS AND LIMITATIONS: We demonstrate that our approach leverages the RCT data for learning and verification, by showing that the DSS suggests treatments that improve the outcome. The results are limited through the small number of patients per treatment; while our ensemble is designed to mitigate this effect, the predictive performance of the methods is affected by the smallness of the data. OUTLOOK: We provide a basis for the establishment of decision supporting routines on treatments that have been tested in RCTs but have not yet been deployed clinically. Practitioners can use our approach to train and validate a DSS on new treatments by simply using the RCT data available to them. More work is needed to strengthen the robustness of the predictors. Since there are no further data available to this purpose, but those already used, the potential of synthetic data generation seems an appropriate alternative.

Smartphones and other mobile devices offer a valu-able opportunity to gather patient-specific health data during everyday life. However, the increasing popularity of mobile health apps demands specialized data analysis methods that can handle the unique, patient-based, time-dependent, and often multivariate data collected by these apps. This work explores the analysis of patient-based mHealth data to develop personalized prediction models. The models can incor-porate data not only from the individual patient, but also from other similar patients using patient-specific neighborhoods. Our approach entails selecting the appropriate data for a particular prediction task and dataset. We also discuss when to utilize data from other patients and offer guidance on selecting similarity functions, models, and model combinations. The approach is illustrated on the case study of tinnitus, a perception of sound without an external source, which can be highly distressing. Its presentation and treatment success are patient-specific. As part of the UNITI project, daily diary data of the patients is collected. Evaluation favored the use of personalized models using patient-specific neighborhoods over a global model using all data, or only using a patient's own data for tinnitus distress prediction.

Ecological momentary assessment (EMA) has been used in many mHealth apps. EMA captures valuable insights into many diseases. Identifying the Granger causal relationships across the EMA variables may contribute to the interpretation of a disease and improve treatment decisions. In our study, we perform a circadian conditional Granger causality analysis on multivariate time series. The analysis was done on EMA data of 270 users of an mHealth app on tinnitus and on their registration data, using the latter to explain the circadian conditional Granger causal relationships. We discovered that some EMA items Granger cause others for more than 8% of the mHealth app users and that these users have answered 8 out of 40 questions at registration differently.

A stream of users' interactions with an mHealth app can be seen as the result of a stochastic process that can be captured by an algorithm that learns over the whole stream. But is it only one process? We investigate to what extend learning for each user separately delivers better predictions than learning one model over the whole stream. Our application scenario is the prediction of Ecological Momentary Assessments (EMA) for an mHealth app (TinnitusTipps) on tinnitus. The data were recorded as part of a pilot study, in which one group of users received non-personalized suggestions (tips) throughout the study, while the other group received tips only during the second half of the study. Our method encompasses user-centric and global stream learning for EMA prediction, combined under a Contextual Multi-Armed Bandit (CMAB) that captures the context of each user group and incorporates the prediction quality of each learner into the reward function. We show that user-centric learning is beneficial for users who contribute many EMA, while a learner over the whole stream is better for users with few EMA.

Mobile Health (mhealth) applications are increasing in popularity, and the collection of disease-specific time series data using Ecological Momentary Assessment (EMA) questionnaires has been shown to help in the creation of personalised predictors for next-step forecasting, which can be crucial in giving preemptive interventions. In this work, we propose a framework that aims to mitigate a common issue in EMA data - that some users contribute a bulk of the data while most users contribute too little. Our proposed framework aims to discover a `useful' neighbourhood of `long' users for `short' ones, by optimising for the error of the user-level predictors for users with little data available for learning. For each user-level predictor, this is done by iteratively adding the next-most-similar long user from a similarity-ordered list as long as the error of the learned model does not increase. This method is compared against a baseline that exploits all available data for the long users, as well as an exhaustive search model that retains only only those users that yield the lowest error predictor. We also explore multiple ways to define similarity, and study the impact of each on the two search strategies on two EMA datasets from an mHealth app `TrackYourDiabetes' - with users from Bulgaria and Spain. Our experiments over the two datasets show a 2.5% and 36.6% improvement respectively for RMSE while using on average 42.8% and 69.9% less data than the baseline method.

Electronic health records (EHR) often include multiple perspectives on a patient's current state of well-being (e.g. vital signs and subjective indicators measured by questionnaires). In this study, we use these perspectives to build phenotypes of chronic tinnitus patients and investigate how these phenotypes are associated with response to treatment. Therefore, we model patients as nodes in a network, where those perspectives are interpreted as layers of a multi-layer network. To identify phenotypes of patients in the network, we implement a community detection algorithm. Some of these communities can be considered as phenotypes if they represent subgroups of patients that are similar according to the investigated perspectives. Furthermore, we analyze the influence of the layers on the final community structure of patients. We then propose a method to add layers given their community structure similarity. Finally, we fit a model, per community, to predict the treatment outcome. In some communities, this prediction outperformed the baseline scenario where the predictor was fitted to all patients.

Abstract Some mHealth apps record user activity continuously and unobtrusively, while other apps rely by nature on user engagement and self-discipline: users are asked to enter data that cannot be assessed otherwise, e.g., on how they feel and what non-measurable symptoms they have. Over time, this leads to substantial differences in the length of the time series of recordings for the different users. In this study, we propose two algorithms for wellbeing-prediction from such time series, and we compare their performance on the users of a pilot study on diabetic patients - with time series length varying between 8 and 87 recordings. Our first approach learns a model from the few users, on which many recordings are available, and applies this model to predict the 2nd, 3rd, and so forth recording of users newly joining the mHealth platform. Our second approach rather exploits the similarity among the first few recordings of newly arriving users. Our results for the first approach indicate that the target variable for users who use the app for long are not predictive for users who use the app only for a short time. Our results for the second approach indicate that few initial recordings suffice to inform the predictive model and improve performance considerably.

Missing data is a common occurrence in the time series domain, for instance due to faulty sensors, server downtime or patients not attending their scheduled appointments. One of the best methods to impute these missing values is Multiple Imputations by Chained Equations (MICE) which has the drawback that it can only model linear relationships among the variables in a multivariate time series. The advancement of deep learning and its ability to model non-linear relationships among variables make it a promising candidate for time series imputation. This work proposes a modified Convolutional Denoising Autoencoder (CDA) based approach to impute multivariate time series data in combination with a preprocessing step that encodes time series data into 2D images using Gramian Angular Summation Field (GASF). We compare our approach against a standard feed-forward Multi Layer Perceptron (MLP) and MICE. All our experiments were performed on 5 UEA MTSC multivariate time series datasets, where 20 to 50% of the data was simulated to be missing completely at random. The CDA model outperforms all the other models in 4 out of 5 datasets and is tied for the best algorithm in the remaining case.

Stream classification algorithms traditionally treat arriving observations as independent. However, in many applications the arriving examples may depend on the "entity" that generated them, e.g. in product reviewing or in the interactions of users with an application server. In this study, we investigate the potential of this dependency by partitioning the original stream of observations into entity-centric substreams and by incorporating entity-specific information into the learning model. We propose a k Nearest Neighbour inspired stream classification approach (kNN), in which the label of an arriving observation is predicted by exploiting knowledge on the observations belonging to this entity and to entities similar to it. For the computation of entity similarity, we consider knowledge about the observations and knowledge about the entity, potentially transferred from another domain. To distinguish between cases where this kind of knowledge transfer is beneficial for stream classification and cases where the knowledge on the entities does not contribute to classifying the observations, we also propose a heuristic approach based on random sampling of substreams using k Random Entities (kRE). Our learning scenario is not fully supervised: after acquiring labels for the initial few observations of each entity, we assume that no additional labels arrive, and attempt to predict the labels of near-future and far-future observations from that initial seed. We report on our findings from three datasets.