Artificial Intelligence in Health

We present MeSHx-Notes, MeSH eXtended for clinical notes, a multi-language web system based on the Django framework to present selected terms in clinical notes. MeSHx-Notes extends Medical Subject Headings (MeSH) terms with Word Embeddings with similar words. Since MeSH is available in 15 languages, MeSHx-Notes is easily extendable by replacing the MeSH thesaurus with the target language (plus the generation of the corresponding WE for the new language). Our version deals with Portuguese and English.

Ensuring sustainable care-giving systems with a focus on human needs and desires is a major challenge. An increasing demand in home health care as well as a limited number of professionals in the labor market have led to a call for efficiency. Thus, managing existing resources has gained relevance. The overall goal is high quality care services, while ensuring economic viability. At the same time, there is a need for modern customer-friendly solutions as well as the consideration of employees’ preferences. To achieve this, adequate methods are needed that take current and future developments into account. Operational management processes in terms of planning and scheduling can be supported by multiagent systems as well as decision support systems using agent-based simulation. The aim of this work is to provide an overview of these solutions in the domain of home health care systems. To this end, we conducted a systematic literature review in which 11 relevant approaches were identified. In addition, these publications were analyzed to identify deficiencies and compared to each other. Because none of the approaches offers a sufficient solution, future work will focus on dynamic distributed scheduling for the control of operational processes which increases efficiency and improves the use of limited resources.

In automated health services based on text and voice interfaces, there is a need to be able to understand what the user is talking about, and what is the attitude of the user towards a subject. Typical machine learning methods for text analysis require a lot of annotated data for the training. This is often a problem in addressing specific and possibly very personal health care needs. In this paper, we propose an active learning algorithm for the training of a text classifier for a conversational therapy application in the area of health behavior change. A new active learning algorithm, Query by Embedded Committee (QBEC), is proposed in the paper. The methods are particularly suitable for the text classification task in a dynamic environment and give a good performance with realistic test data.

The full comprehension of how topics change within psychotherapeutic conversation is key for assessment and therapeutic strategies to adopt by the counselor to the patients. That might enable artificial intelligence (AI) approaches to recommend the most suitable strategy for a new patient. Basically, understanding the topics dynamics of previous cases allows choosing the best therapy to perform for new patients depending on their current conversations.

In this paper we leverage Partially Labeled Dirichlet Allocation with the goal to detect and track topics in real-life psychotherapeutic conversations. On the one hand, the detection of topics allows us identifying the semantic themes of the current therapeutic conversation and predicting topics ad-hoc for each talk-turn between the patient and the counselor. On the other hand, the tracking of topics is key to understand and explore the dynamics of the conversation giving insights and tips on logic and strategy to adopt.

We point out that the entire conversation is structured and modeled according to a sequence of ongoing topics that might propagate through each talk-turn. We present a new method that combines topic modeling and transitions matrices that gives important information to counselors for their therapeutic strategies.

This paper is aimed at formalizing the interplay among a person to be assisted, an assistive agent-based software, and a caregiver. We propose general principles for designing the interplay between a person to be assisted and an agent based on formal argumentation theory to characterize the agent’s reasoning processes. These principles emerge from a novel perspective to understand assistive technology using the concept of zone of proximal development (ZPD) from social sciences. ZPD can be understood as a measurement of activity development, comparing what a person can perform with or without external help. We characterize a rational agent in four ZPD zones: (I) independent activity execution, agent takes no action; (II) \( ZPD_H \): a person supported by another person, agent takes no action; (III) \(ZPD_S\): a person is supported by an agent; and (IV) \( ZPD_{H+S}\): a person is supported by a caregiver and a software agent at the same time. An algorithm was developed for the agent to reason about the actions to be selected in different situations, based on formal argumentation theory for allowing non-monotonic reasoning. The formal models and algorithm were implemented in a prototype system using augmented reality as interface. Future work includes evaluating the principles and algorithm in actual use situations.

The aim of this paper is to introduce a smart mHealth application based on the augmented reality (AR)-paradigm that can support patients with common problems, related to management of their medication. This smart mHealth application is designed and implemented as a medication coach intelligent agent, called Medication Coach Intelligent Agent (MCIA). The MCIA has to manage different types of information such as the medication plan (medication regime) of the patients, medication restrictions, as well as the patient’s preferences and sensor input data from an AR-headset. Considering all this information, the MCIA leads with holistic decisions in order to offer personalized and unobtrusive interventions, in an autonomous way, to the patients. From a long-term perspective, the MCIA should also evaluate its performance over time and adapt in order to improve its interventions with the patients. To show the feasibility of our approach, a proof-of-concept prototype was implemented and evaluated. In this proof-of-concept prototype, the MCIA has been embodied as a smart augmented reality (AR)-mHealth application in the settings of a Microsoft HoloLens. The results show a high potential for using the MCIA in real settings.

Knowledge Management is decisive for clinical decision-making and for delivering better outcomes for patients care. The importance of medical knowledge has been emphasized in the researches to support evidence-based medicine. Currently, cancer is responsible for over 130,000 deaths every year in Brazil. Extensive waiting queues for diagnosis and treatments have become routine. One of the critical success factors in a cancer treatment is the early diagnosis. The reduction of waiting time to start cancer treatment is one of the main issues for improvement of patient’s quality of life and possibilities of cure. This study presents a knowledge-based simulation framework developed at the Brazilian National Cancer Institute (INCA) to reduce patients’ waiting time to start cancer treatment.

The lifted junction tree algorithm (LJT) answers multiple queries efficiently for relational models under uncertainties by building and then reusing a first-order cluster representation. We extend the underling model representation of LJT, which is called parameterised probabilistic model, to calculate a lifted solution to the maximum expected utility (MEU) problem. Specifically, this paper contributes (i) action and utility nodes for parameterised probabilistic models, resulting in parameterised probabilistic decision models and (ii) meuLJT, an algorithm to solve the MEU problem using parameterised probabilistic decision models efficiently, while also being able to answer multiple marginal queries.

This paper presents an overview of the usability engineering process for the development of a personalised clinical decision support system for the management of type 1 diabetes. The tool uses artificial intelligence (AI) techniques to provide insulin bolus dose advice and carbohydrate recommendations that adapt to the individual. We describe the role of human factors and user-centred design in the creation of medical systems that must adhere to international standards. We focus specifically on the formative evaluation stage of this process. The preliminary analysis of data shows promising results.

Accurately determining pain levels in children is difficult, even for trained professionals and parents. Facial activity provides sensitive and specific information about pain, and computer vision algorithms have been developed to automatically detect Facial Action Units (AUs) defined by the Facial Action Coding System (FACS). Our prior work utilized information from computer vision, i.e., automatically detected facial AUs, to develop classifiers to distinguish between pain and no-pain conditions. However, application of pain/no-pain classifiers based on automated AU codings across different environmental domains results in diminished performance. In contrast, classifiers based on manually coded AUs demonstrate reduced environmentally-based variability in performance. In this paper, we train a machine learning model to recognize pain using AUs coded by a computer vision system embedded in a software package called iMotions. We also study the relationship between iMotions (automatically) and human (manually) coded AUs. We find that AUs coded automatically are different from those coded by a human trained in the FACS system, and that the human coder is less sensitive to environmental changes. To improve classification performance in the current work, we applied transfer learning by training another machine learning model to map automated AU codings to a subspace of manual AU codings to enable more robust pain recognition performance when only automatically coded AUs are available for the test data. With this transfer learning method, we improved the Area Under the ROC Curve (AUC) on independent data from new participants in our target domain from 0.67 to 0.72.

Accurately determining pain levels in children is difficult, even for trained professionals and parents. Facial activity and electrodermal activity (EDA) provide rich information about pain, and both have been used in automated pain detection. In this paper, we discuss preliminary steps towards fusing models trained on video and EDA features respectively. We compare fusion models using original video features and those using transferred video features which are less sensitive to environmental changes. We demonstrate the benefit of the fusion and the transferred video features with a special test case involving domain adaptation and improved performance relative to using EDA and video features alone.

In the fight against cancer, cancer registries are an important tool. At the heart of these registries is the data collection and coding process. This process is ruled by complex international standards and numerous best practices, which can easily overwhelm (coding) operators. In this paper, a system assisting operators in the interpretation of best medical coding practices and a short evaluation are presented. By leveraging the arguments used by the coding experts to determine the best coding option, the proposed system answers coding questions from operators and provides a partial explanation for the proposed solution.

Advance care planning, which includes clarifying and documenting goals of care and preferences for future care, is essential for achieving end-of-life care that is consistent with the preferences of dying patients and their families. Physicians document their communication about these preferences as unstructured free text in clinical notes; as a result, routine assessment of this quality indicator is time consuming and costly. In this study, we trained and validated a deep neural network to detect documentation of advanced care planning conversations in clinical notes from electronic health records. We assessed its performance against rigorous manual chart review and rule-based regular expressions. For detecting documentation of patient care preferences at the note level, the algorithm had high performance; F1-score of 92.0 (95% CI, 89.1–95.1), sensitivity of 93.5% (95% CI, 90.0%–98.0%), positive predictive value of 90.5% (95% CI, 86.4%–95.1%) and specificity of 91.0% (95% CI, 86.4%–95.3%) and consistently outperformed regular expression. Deep learning methods offer an efficient and scalable way to improve the visibility of documented serious illness conversations within electronic health record data, helping to better quality of care.

Cancer screening can benefit from individualized decision-making tools that decrease overdiagnosis. The heterogeneity of cancer screening participants advocates the need for more personalized methods. Partially observable Markov decision processes (POMDPs), when defined with an appropriate reward function, can be used to suggest optimal, individualized screening policies. However, determining an appropriate reward function can be challenging. Here, we propose the use of inverse reinforcement learning (IRL) to form rewards functions for lung and breast cancer screening POMDPs. Using experts (physicians) retrospective screening decisions for lung and breast cancer screening, we developed two POMDP models with corresponding reward functions. Specifically, the maximum entropy (MaxEnt) IRL algorithm with an adaptive step size was employed to learn rewards more efficiently; and combined with a multiplicative model to learn state-action pair rewards for a POMDP. The POMDP screening models were evaluated based on their ability to recommend appropriate screening decisions before the diagnosis of cancer. The reward functions learned with the MaxEnt IRL algorithm, when combined with POMDP models in lung and breast cancer screening, demonstrate performance comparable to experts. The Cohen’s Kappa score of agreement between the POMDPs and physicians’ predictions was high in breast cancer and had a decreasing trend in lung cancer.

This empirical study of a complex group of patients with multiple chronic concurrent conditions (diabetes, cardiovascular and kidney diseases) explores the use of deep learning architectures to identify patient segments and contributing factors to 30-day hospital readmissions. We implemented Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) on sequential Electronic Health Records data at the Danderyd Hospital in Stockholm, Sweden. Three distinct sub-types of patient groups were identified: chronic obstructive pulmonary disease, kidney transplant, and paroxysmal ventricular tachycardia. The CNN learned about vector representations of patients, but the RNN was better able to identify and quantify key contributors to readmission such as myocardial infarction and echocardiography. We suggest that vector representations of patients with deep learning should precede predictive modeling of complex patients. The approach also has potential implications for supporting care delivery, care design and clinical decision-making.