This project assumes that planning and management of health systems must be approached as a relevant part of solutions that enable both personalized therapies and the implementation of public health policies. Health AI includes the Machine Learning sub-area which works with computational models that describe relationships between data collected or available in datasets from the health domain. The design of models that represent specific health phenomena associated with different variables is essential to discover innovative knowledge about such phenomena, as well as to support prognoses in the face of different scenarios, thus subsidizing decision making. Descriptive models based on the extraction of causality graphs and explanatory models of high correlation may offer a comprehensive view, not simply correlational but also causal, of the relationship between different variables and different health conditions, and assist in the development of more assertive treatments. Among the descriptive models, we should mention the creation of virtual and/or augmented reality models for the representation of structures present in the patient’s body, constructed from data collected by tomography or magnetic resonance, to be used in the elaboration of personalized treatment strategies or even in the context of robotic surgery. In turn, applied predictive models must be able to translate data from individual patients or populations into diagnoses, prognoses, effective personalized treatments, or even public policies. These models may involve the simulation of scenarios, in order to assist in decision making.
The design of descriptive and predictive models applied to health must adequately deal with a series of technical challenges, including: (1) the exploration of large volumes of heterogeneous data, with different granularities and levels of detail as a representation of the input variables, which significantly increases its complexity; (2) the need for robustness to noise, incompleteness or unbalance of the data, while capturing relevant particularities (exceptions); (3) learning patterns inherent to the data with precision, sensitivity and specificity, based on the ability to transfer learning, or not to depend on supervision; (4) the inference of the parameters of the models in environments of scarcity of data or data with large biases; (5) the construction of integrated models representing different health problems in the same population; (6) the development of new model structures, articulating deterministic or stochastic compartmental models, or models of autonomous agents, superimposed on graphs; (7) the transformation of heterogeneous information into restrictions on epidemic models; (8) the design of adaptation or adjustment processes to mitigate the impact of bias in predictions made by models; (9) the construction of integrated health systems planning and management tools that promote the optimization of the use of resources considering the dimension of the systems’ resilience to emergencies such as epidemics and disasters; (10) the construction of 3D representations in virtual and / or augmented reality for the visualization of data collected by tomography or resonance; and (11) the integration with the explainability framework discussed in Section 4.1.
A relevant part of the modeling effort concerns the representation of the various processes involved in the allocation of resources in health systems. Decision-making assistance systems should be built in order to consider the general health panorama of the population produced by the other models, generating analyzes capable of subsidizing decision-making both in planning and in the logistics of the operation of health networks. It highlights the dimension, which must be explicitly considered, of the resilience of these networks, which must be planned in order to enable the formulation of contingency plans for the treatment of emergencies such as epidemics and disasters. A major challenge that this project aims to address is the development of models capable of producing meaningful interpretations on large volumes of data that would not be treatable, in its raw form, by users. Models are the tools for mapping between available data and their interpretations (diagnoses, prognosis, associations, projections), and between these and therapeutic or management decisions.
The current literature on AI models is a large and growing area. State-of-the-art modeling techniques include black box predictive models such as Deep Learning 122, AutoML 123, hybrid models 124 and multimodal learning 125, compartmental models, Kalman filtering, model ensemble techniques, as well as classic methods (based on in rules, trees, fuzzy, probabilistic) and more recent methods, such as mimic learning, transfer learning 126, knowledge extraction and compression methods, federated learning, contextual models and counterfactual models. All of these approaches may be explored in the context of specific health problems, inspiring adaptations, extensions, combinations and even the development of new techniques more suitable for the particular domain.
The specific research goals or objectives of this research track are: (1) To build models capable of representing the dynamics of relevant health variables in populations, exploring different levels of granularity and techniques, including compartmental models and autonomous agents, taking into account demographic and environmental characteristics. These models must interact with other models aimed at monitoring the health of individuals, allowing both the insertion of epidemiological contexts in the formulation of therapies and the enrichment of data related to public health in the description of the population’s health condition. As a by-product, such models should support the monitoring of communicable diseases, generating early alerts about potential outbreaks and indicating geographical regions at risk; (2) To build self-adaptive models, which allow health professionals to be included in the learning process, interactively, through the labeling of new cases and diagnoses that the system is unable to recognize. The models should evolve incrementally as new cases and innovative knowledge are presented and labeled by specialists, in a synergistic interaction between health professionals (practitioners) and methodologies, models and algorithms. This technology supports the implementation of case-based reasoning in an incremental and automatic way. In this project, new models and learning algorithms will be developed and applied to diagnosis and prognosis. It is expected that the dissemination of the use of the tools to be developed will be relatively simple, allowing for significant gains in scale; (3) To develop studies on the use of robotic actuation devices, for example, the performance of surgeries with the help of robots and intelligent systems and the use of mobile teleoperated robots to act in epidemiological risk environments. In the first case, it is intended to use intelligent control algorithms together with virtual / augmented reality systems and models trained from data from previous surgeries to assist the surgeon. In the second, the intention is to develop and adapt robotic platforms according to the shared control paradigm for carrying out different tasks in environments of epidemiological risk, such as helping patients and health professionals in isolation or disinfecting hospital environments. In these two scenarios, we also intend to research different aspects of Human-Robot Interaction (HRI – Human-Robot Interaction), both in the development of intelligent interfaces for the teleoperation of robots, and in the study of the impact of these technologies on health professionals and patients.
122. Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, et al. A guide to deep learning in healthcare. Nat Med. 2019 Jan;25(1):24–9.
123. Waring J, Lindvall C, Umeton R. Automated machine learning: Review of the state-of-the-art and opportunities for healthcare. Artif Intell Med. 2020 Apr;104:101822.
124. Phillips KA, Trosman JR, Douglas MP. Emergence of Hybrid Models of Genetic Testing Beyond Direct-to-Consumer or Traditional Labs [Internet]. Vol. 321, JAMA. 2019. p. 2403. Available from: http://dx.doi.org/10.1001/jama.2019.5670
125. Palazzo S, Spampinato C, Kavasidis I, Giordano D, Schmidt J, Shah M. Decoding Brain Representations by Multimodal Learning of Neural Activity and Visual Features. IEEE Trans Pattern Anal Mach Intell. 2020;1–1.
126. Deepak S, Ameer PM. Brain tumor classification using deep CNN features via transfer learning. Comput Biol Med. 2019 Aug;111:103345.