RESUMEN
We studied clinical AI-supported decision-making as an example of a high-stakes setting in which explainable AI (XAI) has been proposed as useful (by theoretically providing physicians with context for the AI suggestion and thereby helping them to reject unsafe AI recommendations). Here, we used objective neurobehavioural measures (eye-tracking) to see how physicians respond to XAI with N = 19 ICU physicians in a hospital's clinical simulation suite. Prescription decisions were made both pre- and post-reveal of either a safe or unsafe AI recommendation and four different types of simultaneously presented XAI. We used overt visual attention as a marker for where physician mental attention was directed during the simulations. Unsafe AI recommendations attracted significantly greater attention than safe AI recommendations. However, there was no appreciably higher level of attention placed onto any of the four types of explanation during unsafe AI scenarios (i.e. XAI did not appear to 'rescue' decision-makers). Furthermore, self-reported usefulness of explanations by physicians did not correlate with the level of attention they devoted to the explanations reinforcing the notion that using self-reports alone to evaluate XAI tools misses key aspects of the interaction behaviour between human and machine.
RESUMEN
The influence of AI recommendations on physician behaviour remains poorly characterised. We assess how clinicians' decisions may be influenced by additional information more broadly, and how this influence can be modified by either the source of the information (human peers or AI) and the presence or absence of an AI explanation (XAI, here using simple feature importance). We used a modified between-subjects design where intensive care doctors (N = 86) were presented on a computer for each of 16 trials with a patient case and prompted to prescribe continuous values for two drugs. We used a multi-factorial experimental design with four arms, where each clinician experienced all four arms on different subsets of our 24 patients. The four arms were (i) baseline (control), (ii) peer human clinician scenario showing what doses had been prescribed by other doctors, (iii) AI suggestion and (iv) XAI suggestion. We found that additional information (peer, AI or XAI) had a strong influence on prescriptions (significantly for AI, not so for peers) but simple XAI did not have higher influence than AI alone. There was no correlation between attitudes to AI or clinical experience on the AI-supported decisions and nor was there correlation between what doctors self-reported about how useful they found the XAI and whether the XAI actually influenced their prescriptions. Our findings suggest that the marginal impact of simple XAI was low in this setting and we also cast doubt on the utility of self-reports as a valid metric for assessing XAI in clinical experts.
RESUMEN
PURPOSE: The trajectory of mechanically ventilated patients with coronavirus disease 2019 (COVID-19) is essential for clinical decisions, yet the focus so far has been on admission characteristics without consideration of the dynamic course of the disease in the context of applied therapeutic interventions. METHODS: We included adult patients undergoing invasive mechanical ventilation (IMV) within 48 h of intensive care unit (ICU) admission with complete clinical data until ICU death or discharge. We examined the importance of factors associated with disease progression over the first week, implementation and responsiveness to interventions used in acute respiratory distress syndrome (ARDS), and ICU outcome. We used machine learning (ML) and Explainable Artificial Intelligence (XAI) methods to characterise the evolution of clinical parameters and our ICU data visualisation tool is available as a web-based widget ( https://www.CovidUK.ICU ). RESULTS: Data for 633 adults with COVID-19 who underwent IMV between 01 March 2020 and 31 August 2020 were analysed. Overall mortality was 43.3% and highest with non-resolution of hypoxaemia [60.4% vs17.6%; P < 0.001; median PaO2/FiO2 on the day of death was 12.3(8.9-18.4) kPa] and non-response to proning (69.5% vs.31.1%; P < 0.001). Two ML models using weeklong data demonstrated an increased predictive accuracy for mortality compared to admission data (74.5% and 76.3% vs 60%, respectively). XAI models highlighted the increasing importance, over the first week, of PaO2/FiO2 in predicting mortality. Prone positioning improved oxygenation only in 45% of patients. A higher peak pressure (OR 1.42[1.06-1.91]; P < 0.05), raised respiratory component (OR 1.71[ 1.17-2.5]; P < 0.01) and cardiovascular component (OR 1.36 [1.04-1.75]; P < 0.05) of the sequential organ failure assessment (SOFA) score and raised lactate (OR 1.33 [0.99-1.79]; P = 0.057) immediately prior to application of prone positioning were associated with lack of oxygenation response. Prone positioning was not applied to 76% of patients with moderate hypoxemia and 45% of those with severe hypoxemia and patients who died without receiving proning interventions had more missed opportunities for prone intervention [7 (3-15.5) versus 2 (0-6); P < 0.001]. Despite the severity of gas exchange deficit, most patients received lung-protective ventilation with tidal volumes less than 8 mL/kg and plateau pressures less than 30cmH2O. This was despite systematic errors in measurement of height and derived ideal body weight. CONCLUSIONS: Refractory hypoxaemia remains a major association with mortality, yet evidence based ARDS interventions, in particular prone positioning, were not implemented and had delayed application with an associated reduced responsiveness. Real-time service evaluation techniques offer opportunities to assess the delivery of care and improve protocolised implementation of evidence-based ARDS interventions, which might be associated with improvements in survival.
Asunto(s)
COVID-19 , Respiración Artificial , Adulto , Inteligencia Artificial , Humanos , Posición Prona , SARS-CoV-2 , Reino UnidoRESUMEN
A higher proportion of patients with heart failure have benefitted from a wide and expanding variety of sensor-enabled implantable devices than any other patient group. These patients can now also take advantage of the ever-increasing availability and affordability of consumer electronics. Wearable, on- and near-body sensor technologies, much like implantable devices, generate massive amounts of data. The connectivity of all these devices has created opportunities for pooling data from multiple sensors - so-called interconnectivity - and for artificial intelligence to provide new diagnostic, triage, risk-stratification and disease management insights for the delivery of better, more personalised and cost-effective healthcare. Artificial intelligence is also bringing important and previously inaccessible insights from our conventional cardiac investigations. The aim of this article is to review the convergence of artificial intelligence, sensor technologies and interconnectivity and the way in which this combination is set to change the care of patients with heart failure.
RESUMEN
Dopamine neurons in the substantia nigra pars compacta and ventral tegmental area regulate behaviours such as reward-related learning, and motor control. Dysfunction of these neurons is implicated in Schizophrenia, addiction to drugs, and Parkinson's disease. While some dopamine neurons fire single spikes at regular intervals, others fire irregular single spikes interspersed with bursts. Pharmacological inhibition of calcium-activated potassium (SK) channels increases the variability in their firing pattern, sometimes also increasing the number of spikes fired in bursts, indicating that SK channels play an important role in maintaining dopamine neuron firing regularity and burst firing. However, the exact mechanisms underlying these effects are still unclear. Here, we develop a biophysical model of a dopamine neuron incorporating ion channel stochasticity that enabled the analysis of availability of ion channels in multiple states during spiking. We find that decreased firing regularity is primarily due to a significant decrease in the AHP that in turn resulted in a reduction in the fraction of available voltage-gated sodium channels due to insufficient recovery from inactivation. Our model further predicts that inhibition of SK channels results in a depolarisation of action potential threshold along with an increase in its variability.
Asunto(s)
Potenciales de Acción , Calcio/metabolismo , Neuronas Dopaminérgicas/metabolismo , Mesencéfalo/metabolismo , Canales de Potasio de Pequeña Conductancia Activados por el Calcio/metabolismo , Animales , Agonistas de los Canales de Calcio/farmacología , Dopamina/metabolismo , Neuronas Dopaminérgicas/citología , Neuronas Dopaminérgicas/efectos de los fármacos , Mesencéfalo/citología , Mesencéfalo/efectos de los fármacos , Técnicas de Placa-Clamp , Bloqueadores de los Canales de PotasioRESUMEN
This paper tackles the problem of automatic detection of knee osteoarthritis. A computer system is built that takes as input the body kinetics and produces as output not only an estimation of presence of the knee osteoarthritis, as previously done in the literature, but also the most discriminating parameters along with a set of rules on how this decision was reached. This fills the gap of interpretability between the medical and the engineering approaches. We collected locomotion data from 47 subjects with knee osteoarthritis and 47 healthy subjects. Osteoarthritis subjects were recruited from hospital clinics and GP surgeries, and age and sex matched healthy subjects from the local community. Subjects walked on a walkway equipped with two force plates with piezoelectric 3-component force sensors. Parameters of the vertical, anterior-posterior, and medio-lateral ground reaction forces, such as mean value, push-off time, and slope, were extracted. Then random forest regressors map those parameters via rule induction to the degree of knee osteoarthritis. To boost generalisation ability, a subject-independent protocol is employed. The 5-fold cross-validated accuracy is 72.61%±4.24%. We show that with 3 steps or less a reliable clinical measure can be extracted in a rule-based approach when the dataset is analysed appropriately.
Asunto(s)
Diagnóstico por Computador/métodos , Aprendizaje Automático , Osteoartritis de la Rodilla/diagnóstico , Estudios de Casos y Controles , HumanosRESUMEN
Internal states can profoundly alter the behavior of animals. A quantitative understanding of the behavioral changes upon metabolic challenges is key to a mechanistic dissection of how animals maintain nutritional homeostasis. We used an automated video tracking setup to characterize how amino acid and reproductive states interact to shape exploitation and exploration decisions taken by adult Drosophila melanogaster. We find that these two states have specific effects on the decisions to stop at and leave proteinaceous food patches. Furthermore, the internal nutrient state defines the exploration-exploitation trade-off: nutrient-deprived flies focus on specific patches while satiated flies explore more globally. Finally, we show that olfaction mediates the efficient recognition of yeast as an appropriate protein source in mated females and that octopamine is specifically required to mediate homeostatic postmating responses without affecting internal nutrient sensing. Internal states therefore modulate specific aspects of exploitation and exploration to change nutrient selection.
Asunto(s)
Drosophila melanogaster/fisiología , Animales , Conducta Alimentaria , Preferencias Alimentarias , Homeostasis , Fenómenos Fisiológicos de la Nutrición , Conducta Sexual Animal , Olfato , Grabación en VideoRESUMEN
Functional neuroimaging typically explores how a particular task activates a set of brain regions. Importantly though, the same neural system can be activated by inherently different tasks. To date, there is no approach available that systematically explores whether and how distinct tasks probe the same neural system. Here, we propose and validate an alternative framework, the Automatic Neuroscientist, which turns the standard fMRI approach on its head. We use real-time fMRI in combination with modern machine-learning techniques to automatically design the optimal experiment to evoke a desired target brain state. In this work, we present two proof-of-principle studies involving perceptual stimuli. In both studies optimization algorithms of varying complexity were employed; the first involved a stochastic approximation method while the second incorporated a more sophisticated Bayesian optimization technique. In the first study, we achieved convergence for the hypothesized optimum in 11 out of 14 runs in less than 10 min. Results of the second study showed how our closed-loop framework accurately and with high efficiency estimated the underlying relationship between stimuli and neural responses for each subject in one to two runs: with each run lasting 6.3 min. Moreover, we demonstrate that using only the first run produced a reliable solution at a group-level. Supporting simulation analyses provided evidence on the robustness of the Bayesian optimization approach for scenarios with low contrast-to-noise ratio. This framework is generalizable to numerous applications, ranging from optimizing stimuli in neuroimaging pilot studies to tailoring clinical rehabilitation therapy to patients and can be used with multiple imaging modalities in humans and animals.
Asunto(s)
Mapeo Encefálico/métodos , Imagen por Resonancia Magnética/métodos , Adulto , Algoritmos , Teorema de Bayes , Encéfalo/fisiología , Interfaces Cerebro-Computador , Femenino , Humanos , Procesamiento de Imagen Asistido por Computador/métodos , Aprendizaje Automático , Masculino , Neurociencias/métodosRESUMEN
This study proposes a framework for deconstructing complex walking patterns to create a simple principal component space before checking whether the projection to this space is suitable for identifying changes from the normality. We focus on knee osteoarthritis, the most common knee joint disease and the second leading cause of disability. Knee osteoarthritis affects over 250 million people worldwide. The motivation for projecting the highly dimensional movements to a lower dimensional and simpler space is our belief that motor behaviour can be understood by identifying a simplicity via projection to a low principal component space, which may reflect upon the underlying mechanism. To study this, we recruited 180 subjects, 47 of which reported that they had knee osteoarthritis. They were asked to walk several times along a walkway equipped with two force plates that capture their ground reaction forces along 3 axes, namely vertical, anterior-posterior, and medio-lateral, at 1000 Hz. Data when the subject does not clearly strike the force plate were excluded, leaving 1-3 gait cycles per subject. To examine the complexity of human walking, we applied dimensionality reduction via Probabilistic Principal Component Analysis. The first principal component explains 34% of the variance in the data, whereas over 80% of the variance is explained by 8 principal components or more. This proves the complexity of the underlying structure of the ground reaction forces. To examine if our musculoskeletal system generates movements that are distinguishable between normal and pathological subjects in a low dimensional principal component space, we applied a Bayes classifier. For the tested cross-validated, subject-independent experimental protocol, the classification accuracy equals 82.62%. Also, a novel complexity measure is proposed, which can be used as an objective index to facilitate clinical decision making. This measure proves that knee osteoarthritis subjects exhibit more variability in the two-dimensional principal component space.