Human-centered explainability for life sciences, healthcare, and medical informatics.

Rapid advances in artificial intelligence (AI) and availability of biological, medical, and healthcare data have enabled the development of a wide variety of models. Significant success has been achieved in a wide range of fields, such as genomics, protein folding, disease diagnosis, imaging, and clinical tasks. Although widely used, the inherent opacity of deep AI models has brought criticism from the research field and little adoption in clinical practice. Concurrently, there has been a significant amount of research focused on making such methods more interpretable, reviewed here, but inherent critiques of such explainability in AI (XAI), its requirements, and concerns with fairness/robustness have hampered their real-world adoption. We here discuss how user-driven XAI can be made more useful for different healthcare stakeholders through the definition of three key personas-data scientists, clinical researchers, and clinicians-and present an overview of how different XAI approaches can address their needs. For illustration, we also walk through several research and clinical examples that take advantage of XAI open-source tools, including those that help enhance the explanation of the results through visualization. This perspective thus aims to provide a guidance tool for developing explainability solutions for healthcare by empowering both subject matter experts, providing them with a survey of available tools, and explainability developers, by providing examples of how such methods can influence in practice adoption of solutions.

Blending Knowledge in Deep Recurrent Networks for Adverse Event Prediction at Hospital Discharge.

Deep learning architectures have an extremely high-capacity for modeling complex data in a wide variety of domains. However, these architectures have been limited in their ability to support complex prediction problems using insurance claims data, such as readmission at 30 days, mainly due to data sparsity issue. Consequently, classical machine learning methods, especially those that embed domain knowledge in handcrafted features, are often on par with, and sometimes outperform, deep learning approaches. In this paper, we illustrate how the potential of deep learning can be achieved by blending domain knowledge within deep learning architectures to predict adverse events at hospital discharge, including readmissions. More specifically, we introduce a learning architecture that fuses a representation of patient data computed by a self-attention based recurrent neural network, with clinically relevant features. We conduct extensive experiments on a large claims dataset and show that the blended method outperforms the standard machine learning approaches.

Disease network delineates the disease progression profile of cardiovascular diseases.

OBJECTIVE: As Electronic Health Records (EHR) data accumulated explosively in recent years, the tremendous amount of patient clinical data provided opportunities to discover real world evidence. In this study, a graphical disease network, named progressive cardiovascular disease network (progCDN), was built to delineate the progression profiles of cardiovascular diseases (CVD). MATERIALS AND METHODS: The EHR data of 14.3 million patients with CVD diagnoses were collected for building disease network and further analysis. We applied a new designed method, progression rates (PR), to calculate the progression relationship among different diagnoses. Based on the disease network outcome, 23 disease progression pair were selected to screen for salient features. RESULTS: The network depicted the dominant diseases in CVD development, such as the heart failure and coronary arteriosclerosis. Novel progression relationships were also discovered, such as the progression path from long QT syndrome to major depression. In addition, three age-group progCDNs identified a series of age-associated disease progression paths and important successor diseases with age bias. Furthermore, a list of important features with sufficient abundance and high correlation was extracted for building disease risk models. DISCUSSION: The PR method designed for identifying the progression relationship could be widely applied in any EHR database due to its flexibility and robust functionality. Meanwhile, researchers could use the progCDN network to validate or explore novel disease relationships in real world data. CONCLUSION: The first-time interrogation of such a huge CVD patients cohort enabled us to explore the general and age-specific disease progression patterns in CVD development.

Impact of Clinical and Genomic Factors on COVID-19 Disease Severity.

To date, there have been 180 million confirmed cases of COVID-19, with more than 3.8 million deaths, reported to WHO worldwide. In this paper we address the problem of understanding the host genome's influence, in concert with clinical variables, on the severity of COVID-19 manifestation in the patient. Leveraging positive-unlabeled machine learning algorithms coupled with RubricOE, a state-of-the-art genomic analysis framework, on UK BioBank data we extract novel insights on the complex interplay. The algorithm is also sensitive enough to detect the changing influence of the emergent B.1.1.7 SARS-CoV-2 (alpha) variant on disease severity, and, changing treatment protocols. The genomic component also implicates biological pathways that can help in understanding the disease etiology. Our work demonstrates that it is possible to build a robust and sensitive model despite significant bias, noise and incompleteness in both clinical and genomic data by a careful interleaving of clinical and genomic methodologies.

Finding Causal Mechanistic Drug-Drug Interactions from Observational Data.

Many adverse drug reactions (ADRs) are caused by drug-drug interactions (DDIs), meaning they arise from concurrent use of multiple medications. Detecting DDIs using observational data has at least three major challenges: (1) The number of potential DDIs is astronomical; (2) Associations between drugs and ADRs may not be causal due to observed or unobserved confounding; and (3) Frequently co-prescribed drug pairs that each independently cause an ADR do not necessarily causally interact, where causal interaction means that at least some patients would only experience the ADR if they take both drugs. We address (1) through data mining algorithms pre-filtering potential interactions, and (2) and (3) by fitting causal interaction models adjusting for observed confounders and conducting sensitivity analyses for unobserved confounding. We rank candidate DDIs robust to unobserved confounding more likely to be real. Our rigorous approach produces far fewer false positives than past applications that ignored (2) and (3).

Is Deep Reinforcement Learning Ready for Practical Applications in Healthcare? A Sensitivity Analysis of Duel-DDQN for Hemodynamic Management in Sepsis Patients.

The potential of Reinforcement Learning (RL) has been demonstrated through successful applications to games such as Go and Atari. However, while it is straightforward to evaluate the performance of an RL algorithm in a game setting by simply using it to play the game, evaluation is a major challenge in clinical settings where it could be unsafe to follow RL policies in practice. Thus, understanding sensitivity of RL policies to the host of decisions made during implementation is an important step toward building the type of trust in RL required for eventual clinical uptake. In this work, we perform a sensitivity analysis on a state-of-the-art RL algorithm (Dueling Double Deep Q-Networks) applied to hemodynamic stabilization treatment strategies for septic patients in the ICU. We consider sensitivity of learned policies to input features, embedding model architecture, time discretization, reward function, and random seeds. We find that varying these settings can significantly impact learned policies, which suggests a need for caution when interpreting RL agent output.

Real-Time Processing of Continuous Physiological Signals in a Neurocritical Care Unit on a Stream Data Analytics Platform.

Continuous high-volume and high-frequency brain signals such as intracranial pressure (ICP) and electroencephalographic (EEG) waveforms are commonly collected by bedside monitors in neurocritical care. While such signals often carry early signs of neurological deterioration, detecting these signs in real time with conventional data processing methods mainly designed for retrospective analysis has been extremely challenging. Such methods are not designed to handle the large volumes of waveform data produced by bedside monitors. In this pilot study, we address this challenge by building a prototype system using the IBM InfoSphere Streams platform, a scalable stream computing platform, to detect unstable ICP dynamics in real time. The system continuously receives electrocardiographic and ICP signals and analyzes ICP pulse morphology looking for deviations from a steady state. We also designed a Web interface to display in real time the result of this analysis in a Web browser. With this interface, physicians are able to ubiquitously check on the status of their patients and gain direct insight into and interpretation of the patient's state in real time. The prototype system has been successfully tested prospectively on live hospitalized patients.

Heart rate variability for preclinical detection of secondary complications after subarachnoid hemorrhage.

BACKGROUND: We sought to determine if monitoring heart rate variability (HRV) would enable preclinical detection of secondary complications after subarachnoid hemorrhage (SAH). METHODS: We studied 236 SAH patients admitted within the first 48 h of bleed onset, discharged after SAH day 5, and had continuous electrocardiogram records available. The diagnosis and date of onset of infections and DCI events were prospectively adjudicated and documented by the clinical team. Continuous ECG was collected at 240 Hz using a high-resolution data acquisition system. The Tompkins-Hamilton algorithm was used to identify R-R intervals excluding ectopic and abnormal beats. Time, frequency, and regularity domain calculations of HRV were generated over the first 48 h of ICU admission and 24 h prior to the onset of each patient's first complication, or SAH day 6 for control patients. Clinical prediction rules to identify infection and DCI events were developed using bootstrap aggregation and cost-sensitive meta-classifiers. RESULTS: The combined infection and DCI model predicted events 24 h prior to clinical onset with high sensitivity (87 %) and moderate specificity (66 %), and was more sensitive than models that predicted either infection or DCI. Models including clinical and HRV variables together substantially improved diagnostic accuracy (AUC 0.83) compared to models with only HRV variables (AUC 0.61). CONCLUSIONS: Changes in HRV after SAH reflect both delayed ischemic and infectious complications. Incorporation of concurrent disease severity measures substantially improves prediction compared to using HRV alone. Further research is needed to refine and prospectively evaluate real-time bedside HRV monitoring after SAH.

Real-time multidimensional temporal analysis of complex high volume physiological data streams in the neonatal intensive care unit.

The intensive care of immature preterm infants is a challenging, dynamic clinical task that is complicated because these infants frequently develop a range of comorbidities as they grow and develop after their premature birth. Earliest reliable condition onset detection is a goal within this setting and high frequency physiological analysis is showing potential new pathophysiological indicators for earlier onset detection of several conditions. To realise this, a platform for multi-stream, multi-condition, multi-feature risk scoring is required. In this paper we demonstrate our multi-stream online analytics approach for condition onset detection and demonstrate a user interface approach for patient state that can be available in real-time to support condition risk scoring.

A discriminative approach to EEG seizure detection.

Seizures are abnormal sudden discharges in the brain with signatures represented in electroencephalograms (EEG). The efficacy of the application of speech processing techniques to discriminate between seizure and non-seizure states in EEGs is reported. The approach accounts for the challenges of unbalanced datasets (seizure and non-seizure), while also showing a system capable of real-time seizure detection. The Minimum Classification Error (MCE) algorithm, which is a discriminative learning algorithm with wide-use in speech processing, is applied and compared with conventional classification techniques that have already been applied to the discrimination between seizure and non-seizure states in the literature. The system is evaluated on 22 pediatric patients multi-channel EEG recordings. Experimental results show that the application of speech processing techniques and MCE compare favorably with conventional classification techniques in terms of classification performance, while requiring less computational overhead. The results strongly suggests the possibility of deploying the designed system at the bedside.

Real-time prognosis of ICU physiological data streams.

This paper presents a system capable of predicting in real-time the evolution of Intensive Care Unit (ICU) physiological patient data streams. It leverages a state of the art stream computing platform to host analytics capable of making such prognosis in real time. The focus is on online algorithms that do not require a training phase. We use Fading-Memory Polynomial filters [8] on the frequency domain to predict windows of ICU data streams. We report on both the system and the performance of this approach when applied to traces of more than 1500 ICU patients obtained from the MIMIC-II database [1].

Real-time analysis for intensive care: development and deployment of the artemis analytic system.

The lives of many thousands of children born premature or ill at term around the world have been saved by those who work within neonatal intensive care units (NICUs). Modern-day neonatologists, together with nursing staff and other specialists within this domain, enjoy modern technologies for activities such as financial transactions, online purchasing, music, and video on demand. Yet, when they move into their workspace, in many cases, they are supported by nearly the same technology they used 20 years ago. Medical devices provide visual displays of vital signs through physiological streams such as electrocardiogram (ECG), heart rate, blood oxygen saturation (SpO(2)), and respiratory rate. Electronic health record initiatives around the world provide an environment for the electronic management of medical records, but they fail to support the high-frequency interpretation of streaming physiological data. We have taken a collaborative research approach to address this need to provide a flexible platform for the real-time online analysis of patients' data streams to detect medically significant conditions that precede the onset of medical complications. The platform supports automated or clinician-driven knowledge discovery to discover new relationships between physiological data stream events and latent medical conditions as well as to refine existing analytics. Patients benefit from the system because earlier detection of signs of the medical conditions may lead to earlier intervention that may potentially lead to improved patient outcomes and reduced length of stays. The clinician benefits from a decision support tool that provides insight into multiple streams of data that are too voluminous to assess with traditional methods. The remainder of this article summarizes the strengths of our research collaboration and the resulting environment known as Artemis, which is currently being piloted within the NICU of The Hospital for Sick Children (SickKids) in Toronto, Ontario, Canada. Although the discussion in this article focuses on a NICU, the technologies can be applied to any intensive care environment.

Predicting Patient's Trajectory of Physiological Data using Temporal Trends in Similar Patients: A System for Near-Term Prognostics.

Providing near-term prognostic insight to clinicians helps them to better assess the near-term impact of their decisions and potential impending events affecting the patient. In this work, we present a novel system, which leverages inter-patient similarity for retrieving patients who display similar trends in their physiological time-series data. Data from the retrieved patient cohort is then used to project patient data into the future to provide insights for the query patient. The proposed approach and system were tested using the MIMIC II database, which consists of physiological waveforms, and accompanying clinical data obtained for ICU patients. In the experiments we report the effectiveness of the inter-patient similarity measure and the accuracy of the projection of patients' data. We also discuss the visual interface that conveys the near-term prognostic decision support to the user.