Explainable deep learning in healthcare: A methodological survey from an attribution view.

The increasing availability of large collections of electronic health record (EHR) data and unprecedented technical advances in deep learning (DL) have sparked a surge of research interest in developing DL based clinical decision support systems for diagnosis, prognosis, and treatment. Despite the recognition of the value of deep learning in healthcare, impediments to further adoption in real healthcare settings remain due to the black-box nature of DL. Therefore, there is an emerging need for interpretable DL, which allows end users to evaluate the model decision making to know whether to accept or reject predictions and recommendations before an action is taken. In this review, we focus on the interpretability of the DL models in healthcare. We start by introducing the methods for interpretability in depth and comprehensively as a methodological reference for future researchers or clinical practitioners in this field. Besides the methods' details, we also include a discussion of advantages and disadvantages of these methods and which scenarios each of them is suitable for, so that interested readers can know how to compare and choose among them for use. Moreover, we discuss how these methods, originally developed for solving general-domain problems, have been adapted and applied to healthcare problems and how they can help physicians better understand these data-driven technologies. Overall, we hope this survey can help researchers and practitioners in both artificial intelligence and clinical fields understand what methods we have for enhancing the interpretability of their DL models and choose the optimal one accordingly. This article is categorized under: Cancer > Computational Models.

Non-Cell-Autonomous Regulation of Optic Nerve Regeneration by Amacrine Cells.

Visual information is conveyed from the eye to the brain through the axons of retinal ganglion cells (RGCs) that course through the optic nerve and synapse onto neurons in multiple subcortical visual relay areas. RGCs cannot regenerate their axons once they are damaged, similar to most mature neurons in the central nervous system (CNS), and soon undergo cell death. These phenomena of neurodegeneration and regenerative failure are widely viewed as being determined by cell-intrinsic mechanisms within RGCs or to be influenced by the extracellular environment, including glial or inflammatory cells. However, a new concept is emerging that the death or survival of RGCs and their ability to regenerate axons are also influenced by the complex circuitry of the retina and that the activation of a multicellular signaling cascade involving changes in inhibitory interneurons - the amacrine cells (AC) - contributes to the fate of RGCs. Here, we review our current understanding of the role that interneurons play in cell survival and axon regeneration after optic nerve injury.

Dual-Purpose SERS Sensor for Selective Determination of Polycyclic Aromatic Compounds via Electron Donor-Acceptor Traps.

Toxic, carcinogenic, and mutagenic properties of polycyclic aromatic hydrocarbons (PAHs) and environmental pollution caused by polycyclic aromatic sulfur heterocycles (PASHs) postulate the importance of their selective and sensitive determination in environmental and oil fuel samples. Surface-enhanced Raman spectroscopy (SERS) opens up an avenue toward multiplex analysis of complex mixtures, however not every molecule gives high enhancement factors and, thus, cannot be reliably detected via SERS. However, the sensitivity can be drastically increased by additional resonant enhancement as a result of the analyte absorption band overlapping with the surface plasmon band of nanoparticles (NPs) and the laser excitation wavelength. Using this idea, we developed a dual-purpose SERS sensor based on trapping the target PAHs and PASHs into colored charge-transfer complexes (CTCs) with selected organic π-acceptor molecules on the surface of AgNPs. Studying, computing, and then comparing stability constants of the formed CTC served as a powerful explanation and prediction tool for a wise choice of π-acceptor indicator systems for the further silver surface modification. Moreover, we show that CTC formation can be effectively utilized for increasing both selectivity and sensitivity by simple liquid-liquid extraction prior to SERS measurements. For the first time, the dual-purpose SERS sensor allowed determination of two different classes of polycyclic aromatic fuel components down to 10 nM concentration, lower than that restricted by the ASTM regulation, and demonstrated multi-purpose capabilities of the developed approach.

Visual recovery following optic nerve crush in male and female wild-type and TRIF-deficient mice.

BACKGROUND: There is growing evidence that the TIR-domain-containing adapter-inducing interferon-ß (TRIF) pathway is implicated in the modulation of neuroinflammation following injuries to the brain and retina. After exposure to injury or to excitotoxic pathogens, toll-like receptors (TLR) activate the innate immune system signaling cascade and stimulate the release of inflammatory cytokines. Inhibition of the TLR4 receptor has been shown to enhance retinal ganglion cell (RGC) survival in optic nerve crush (ONC) and in ischemic injury to other parts of the brain. OBJECTIVE: Based on this evidence, we tested the hypothesis that mice with the TRIF gene knocked out (TKO) will demonstrate decreased inflammatory responses and greater functional recovery after ONC. METHODS: Four experimental groups -TKO ONC (12 males and 8 females), WT ONC (10 males and 8 females), TKO sham (9 males and 5 females), and WT sham (7 males and 5 females) -were used as subjects. Visual evoked potentials (VEP) were recorded in the left and right primary visual cortices and optomotor response were assessed in all mice at 14, 30, and 80 days after ONC. GFAP and Iba-1 were used as markers for astrocytes and microglial cells respectively at 7 days after ONC, along with NF-kB to measure inflammatory effects downstream of TRIF activation; RMPBS marker was used to visualize RGC survival and GAP-43 was used as a marker of regenerating optic nerve axons at 30 days after ONC. RESULTS: We found reduced inflammatory response in the retina at 7 days post-ONC, less RGC loss and greater axonal regeneration 30 days post-ONC, and better recovery of visual function 80 days post-ONC in TKO mice compared to WT mice. CONCLUSIONS: Our study showed that the TRIF pathway is involved in post-ONC inflammatory response and gliosis and that deletion of TRIF induces better RGC survival and regeneration and better functional recovery in mice. Our results suggest the TRIF pathway as a potential therapeutic target for reducing the inflammatory damage caused by nervous system injury.

Vision modulation, plasticity and restoration using non-invasive brain stimulation - An IFCN-sponsored review.

The visual system has one of the most complex structures of all sensory systems and is perhaps the most important sense for everyday life. Its functional organization was extensively studied for decades in animal and humans, for example by correlating circumscribed anatomical lesions in patients with the resulting visual dysfunction. During the past two decades, significant achievements were accomplished in characterizing and modulating visual information processing using non-invasive stimulation techniques of the normal and damaged human eye and brain. Techniques include transcranial magnetic stimulation (TMS) and low intensity electric stimulation using either direct or alternating currents applied transcranially (tDCS or tACS) near or above the visual cortex, or alternating currents applied transorbitally (trACS). In the case of transorbital stimulation of the visual system the electrodes are attached near the eye, to the eyelids (transpalpebral electrical stimulation - TPES) or the cornea (tanscorneal electrical stimulation TcES). Here, we summarize the state-of-the-art of visual system magnetic and electric stimulation as a method to modulate normal vision, induce brain plasticity, and to restore visual functions in patients. We review this field's history, models of current flow paths in the eye and brain, neurophysiological principles (e.g. entrainment and after-effects), the effects on vision in normal subjects and the clinical impact on plasticity and vision restoration in patients with low vision, with a particular focus on "off-line" or "after-effects". With regard to the therapeutic possibilities, ACS was demonstrated to be effective in patients affected by glaucoma and optic neuropathy, while tDCS and random noise stimulation (tRNS) are most promising for the treatment of amblyopia, hemianopia and myopia. In addition, rTMS applied above the occipital area is a promising approach to treat migraine, neglect and hemianopia. Although the response to these treatment options is better than to sham stimulation in double blinded clinical studies, the clinical efficacy is still rather variable and a proportion of patients do not respond. It is therefore imperative to better understand the mechanisms of action to be able to optimize treatment protocols possibly through personalization of brain stimulation protocols. By identifying the current opportunities and challenges in the field, we hope to provide insights to help improve neuromodulation protocols to restore visual function in patients with visual system damage.

Optic nerve regeneration: A long view.

The optic nerve conveys information about the outside world from the retina to multiple subcortical relay centers. Until recently, the optic nerve was widely believed to be incapable of re-growing if injured, with dire consequences for victims of traumatic, ischemic, or neurodegenerative diseases of this pathway. Over the past 10-20 years, research from our lab and others has made considerable progress in defining factors that normally suppress axon regeneration and the ability of retinal ganglion cells, the projection neurons of the retina, to survive after nerve injury. Here we describe research from our lab on the role of inflammation-derived growth factors, suppression of inter-cellular signals among diverse retinal cell types, and combinatorial therapies, along with related studies from other labs, that enable animals with optic nerve injury to regenerate damaged retinal axons back to the brain. These studies raise the possibility that vision might one day be restored to people with optic nerve damage.

Negation Scope Detection in Clinical Notes and Scientific Abstracts: A Feature-enriched LSTM-based Approach.

Electronic Health Records contain a wealth of clinical information that can potentially be used for a variety of clinical tasks. Clinical narratives contain information about the existence or absence of medical conditions as well as clinical findings. It is essential to be able to distinguish between the two since the negated events and the non-negated events often have very different prognostic value. In this paper, we present a feature-enriched neural network-based model for negation scope detection in biomedical texts. The system achieves a robust high performance on two different types of texts, scientific abstracts, and radiology reports, achieving the new state-of-the-art result without requiring the availability of gold cue information for negation scope detection task on the scientific abstracts part of BioScope1 corpus and competitive result on the radiology report corpus.

Progesterone Treatment Attenuates Glycolytic Metabolism and Induces Senescence in Glioblastoma.

We examined the effect of progesterone treatments on glycolytic metabolism and senescence as possible mechanisms in controlling the growth of glioblastoma multiforme (GBM). In an orthotopic mouse model, after tumor establishment, athymic nude mice received treatment with progesterone or vehicle for 40 days. Compared to controls, high-dose progesterone administration produced a significant reduction in tumor size (~47%) and an increased survival rate (~43%) without any demonstrable toxicity to peripheral organs (liver, kidney). This was accompanied by a significant improvement in spontaneous locomotor activity and reduced anxiety-like behavior. In a follow-up in vitro study of U87MG-luc, U87dEGFR and U118MG tumor cells, we observed that high-dose progesterone inhibited expression of Glut1, which facilitated glucose transport into the cytoplasm; glyceraldehyde 3-phosphate dehydrogenase (GAPDH; a glycolysis enzyme); ATP levels; and cytoplasmic FoxO1 and Phospho-FoxO1, both of which control glycolytic metabolism through upstream PI3K/Akt/mTOR signaling in GBM. In addition, progesterone administration attenuated EGFR/PI3K/Akt/mTOR signaling, which is highly activated in grade IV GBM. High-dose progesterone also induced senescence in GBM as evidenced by changes in cell morphology and ß-galactocidase accumulation. In conclusion, progesterone inhibits the modulators of glycolytic metabolism and induces premature senescence in GBM cells and this can help to reduce/slow tumor progression.

Neurosteroid allopregnanolone reduces ipsilateral visual cortex potentiation following unilateral optic nerve injury.

In adult mice with unilateral optic nerve crush injury (ONC), we studied visual response plasticity in the visual cortex following stimulation with sinusoidal grating. We examined visually evoked potentials (VEP) in the primary visual cortex ipsilateral and contralateral to the crushed nerve. We found that unilateral ONC induces enhancement of visual response on the side ipsilateral to the injury that is evoked by visual stimulation to the intact eye. This enhancement was associated with supranormal spatial frequency thresholds in the intact eye when tested using optomotor response. To probe whether injury-induced disinhibition caused the potentiation, we treated animals with the neurosteroid allopregnanolone, a potent agonist of the GABAA receptor, one hour after crush and on post-injury days 3, 8, 13, and 18. Allopregnanolone diminished enhancement of the VEP and this effect was associated with the upregulated synthesis of the δ-subunit of the GABAA receptor. Our study shows a new aspect of experience-dependent plasticity following unilateral ONC. This hyper-activity in the ipsilateral visual cortex is prevented by upregulation of GABA inhibition with allopregnanolone. Our findings suggest the therapeutic potential of allopregnanolone for modulation of plasticity in certain eye and brain disorders and a possible role for disinhibition in ipsilateral hyper-activity following unilateral ONC.

Non-invasive electrical brain stimulation: from acute to late-stage treatment of central nervous system damage.

Non-invasive brain current stimulation (NIBS) is a promising and versatile tool for inducing neuroplasticity, protection and functional rehabilitation of damaged neuronal systems. It is technically simple, requires no surgery, and has significant beneficial effects. However, there are various technical approaches for NIBS which influence neuronal networks in significantly different ways. Transcranial direct current stimulation (tDCS), alternating current stimulation (ACS) and repetitive transcranial magnetic stimulation (rTMS) all have been applied to modulate brain activity in animal experiments under normal and pathological conditions. Also clinical trials have shown that tDCS, rTMS and ACS induce significant behavioural effects and can - depending on the parameters chosen - enhance or decrease brain excitability and influence performance and learning as well as rehabilitation and protective mechanisms. The diverse phaenomena and partially opposing effects of NIBS are not yet fully understood and mechanisms of action need to be explored further in order to select appropriate parameters for a given task, such as current type and strength, timing, distribution of current densities and electrode position. In this review, we will discuss the various parameters which need to be considered when designing a NIBS protocol and will put them into context with the envisaged applications in experimental neurobiology and medicine such as vision restoration, motor rehabilitation and cognitive enhancement.

Cholinergic Potentiation of Restoration of Visual Function after Optic Nerve Damage in Rats.

Enhancing cortical plasticity and brain connectivity may improve residual vision following a visual impairment. Since acetylcholine plays an important role in attention and neuronal plasticity, we explored whether potentiation of the cholinergic transmission has an effect on the visual function restoration. To this end, we evaluated for 4 weeks the effect of the acetylcholinesterase inhibitor donepezil on brightness discrimination, visually evoked potentials, and visual cortex reactivity after a bilateral and partial optic nerve crush in adult rats. Donepezil administration enhanced brightness discrimination capacity after optic nerve crush compared to nontreated animals. The visually evoked activation of the primary visual cortex was not restored, as measured by evoked potentials, but the cortical neuronal activity measured by thallium autometallography was not significantly affected four weeks after the optic nerve crush. Altogether, the results suggest a role of the cholinergic system in postlesion cortical plasticity. This finding agrees with the view that restoration of visual function may involve mechanisms beyond the area of primary damage and opens a new perspective for improving visual rehabilitation in humans.

Stress primes microglial polarization after global ischemia: Therapeutic potential of progesterone.

Despite the fact that stress is associated with increased risk of stroke and worsened outcome, most preclinical studies have ignored this comorbid factor, especially in the context of testing neuroprotective treatments. Preclinical research suggests that stress primes microglia, resulting in an enhanced reactivity to a subsequent insult and potentially increasing vulnerability to stroke. Ischemia-induced activated microglia can be polarized into a harmful phenotype, M1, which produces pro-inflammatory cytokines, or a protective phenotype, M2, which releases anti-inflammatory cytokines and neurotrophic factors. Selective modulation of microglial polarization by inhibiting M1 or stimulating M2 may be a potential therapeutic strategy for treating cerebral ischemia. Our laboratory and others have shown progesterone to be neuroprotective against ischemic stroke in rodents, but it is not known whether it will be as effective under a comorbid condition of chronic stress. Here we evaluated the neuroprotective effect of progesterone on the inflammatory response in the hippocampus after exposure to stress followed by global ischemia. We focused on the effects of microglial M1/M2 polarization and pro- and anti-inflammatory mediators in stressed ischemic animals. Male Sprague-Dawley rats were exposed to 8 consecutive days of social defeat stress and then subjected to global ischemia or sham surgery. The rats received intraperitoneal injections of progesterone (8mg/kg) or vehicle at 2h post-ischemia followed by subcutaneous injections at 6h and once every 24h post-injury for 7days. The animals were killed at 7 and 14days post-ischemia, and brains were removed and processed to assess outcome measures using histological, immunohistochemical and molecular biology techniques. Pre-ischemic stress (1) exacerbated neuronal loss and neurodegeneration as well as microglial activation in the selectively vulnerable CA1 hippocampal region, (2) dysregulated microglial polarization, leading to upregulation of both M1 and M2 phenotype markers, (3) increased pro-inflammatory cytokine expression, and (4) reduced anti-inflammatory cytokine and neurotrophic factor expression in the ischemic hippocampus. Treatment with progesterone significantly attenuated stress-induced microglia priming by modulating polarized microglia and the inflammatory environment in the hippocampus, the area most vulnerable to ischemic injury. Our findings can be taken to suggest that progesterone holds potential as a candidate for clinical testing in ischemic stroke where high stress may be a contributing factor.

Electrical brain stimulation induces dendritic stripping but improves survival of silent neurons after optic nerve damage.

Repetitive transorbital alternating current stimulation (rtACS) improves vision in patients with chronic visual impairments and an acute treatment increased survival of retinal neurons after optic nerve crush (ONC) in rodent models of visual system injury. However, despite this protection no functional recovery could be detected in rats, which was interpreted as evidence of "silent survivor" cells. We now analysed the mechanisms underlying this "silent survival" effect. Using in vivo microscopy of the retina we investigated the survival and morphology of fluorescent neurons before and after ONC in animals receiving rtACS or sham treatment. One week after the crush, more neurons survived in the rtACS-treated group compared to sham-treated controls. In vivo imaging further revealed that in the initial post-ONC period, rtACS induced dendritic pruning in surviving neurons. In contrast, dendrites in untreated retinae degenerated slowly after the axonal trauma and neurons died. The complete loss of visual evoked potentials supports the hypothesis that cell signalling is abolished in the surviving neurons. Despite this evidence of "silencing", intracellular free calcium imaging showed that the cells were still viable. We propose that early after trauma, complete dendritic stripping following rtACS protects neurons from excitotoxic cell death by silencing them.

Preclinical model of transcorneal alternating current stimulation in freely moving rats.

PURPOSE: Transcorneal alternating current stimulation (tACS) has become a promising tool to modulate brain functions and treat visual diseases. To understand the mechanisms of action a suitable animal model is required. However, because existing animal models employ narcosis, which interferes with brain oscillations and stimulation effects, we developed an experimental setup where current stimulation via the eye and flicker light stimulation can be applied while simultaneously recording local field potentials in awake rats. METHOD: tACS was applied in freely-moving rats (N = 24) which had wires implanted under their upper eye lids. Field potential recordings were made in visual cortex and superior colliculus. To measure visual evoked responses, rats were exposed to flicker-light using LEDs positioned in headset spectacles. RESULTS: Corneal electrodes and recording assemblies were reliably operating and well tolerated for at least 4 weeks. Transcorneal stimulation without narcosis did not induce any adverse reactions. Stable head stages allowed repetitive and long-lasting recordings of visual and electrically evoked potentials in freely moving animals. Shape and latencies of electrically evoked responses measured in the superior colliculus and visual cortex indicate that specific physiological responses could be recorded after tACS. CONCLUSIONS: Our setup allows the stimulation of the visual system in unanaesthetised rodents with flicker light and transcorneally applied current travelling along the physiological signalling pathway. This methodology provides the experimental basis for further studies of recovery and restoration of vision.

Retinal origin of electrically evoked potentials in response to transcorneal alternating current stimulation in the rat.

PURPOSE: Little is known about the physiological mechanisms underlying the reported therapeutic effects of transorbital alternating current stimulation (ACS) in vision restoration, or the origin of the recorded electrically evoked potentials (EEPs) during such stimulation. We examined the issue of EEP origin and electrode configuration for transorbital ACS and characterized the physiological responses to CS in different structures of the visual system. METHODS: We recorded visually evoked potentials (VEPs) and EEPs from the rat retina, visual thalamus, tectum, and visual cortex. The VEPs were evoked by light flashes and EEPs were evoked by electric stimuli delivered by two electrodes placed either together on the same eye or on the eyeball and in the neck. Electrically evoked potentials and VEPs were recorded before and after bilateral intraorbital injections of tetrodotoxin that blocked retinal ganglion cell activity. RESULTS: Tetrodotoxin abolished VEPs at all levels in the visual pathway, confirming successful blockage of ganglion cell activity. Tetrodotoxin also abolished EEPs and this effect was independent of the stimulating electrode configurations. CONCLUSIONS: Transorbital electrically evoked responses in the visual pathway, irrespective of reference electrode placement, are initiated by activation of the retina and not by passive conductance and direct activation of neurons in other visual structures. Thus, placement of stimulating electrodes exclusively around the eyeball may be sufficient to achieve therapeutic effects.

Transcorneal alternating current stimulation after severe axon damage in rats results in "long-term silent survivor" neurons.

Transcorneal alternating current stimulation (tACS) was proposed to decrease acute death of retinal ganglion cells after optic nerve transection in rats, but it is not known if cell survival is long-term and associated with functional restoration. We therefore evaluated the effects of tACS in a rat model of optic nerve crush using anatomical, electrophysiological and behavioural measures. Rats were trained in a brightness discrimination visual task and the retinal ganglion cell number was quantified with in vivo confocal neuroimaging. Thereafter, severe optic nerve crush or sham crush was performed and rats were treated under anaesthesia either with tACS or sham stimulation immediately after the lesion and on day 3, 7, 11, 15, 19 and 23. Brightness discrimination was evaluated for 6 weeks and retinal ganglion cells were counted in vivo on post-crush days 7 and 28. In additional rats we studied the influence of tACS on bioelectrical activity. On post-lesion day 28, the tACS-treated group showed a neuronal survival of 28.2% which was significantly greater than in sham operates (8.6%). All animals with optic nerve crush were significantly impaired in brightness discrimination and did not recover performance, irrespective to which group they belonged. In accordance with this, there was no significant influence of the stimulation on EEG power spectra. In conclusion, tACS induced long-term neuronal protection from delayed retrograde cell death, but in this case of severe axonal damage tACS did not influence functional restoration and EEG signals recorded over the visual cortex.

Transcorneal alternating current stimulation induces EEG "aftereffects" only in rats with an intact visual system but not after severe optic nerve damage.

Noninvasive alternating current stimulation can induce vision restoration in patients with chronic optic nerve damage and results in electroencephalogram (EEG) aftereffects. To better understand the mechanisms of action, we studied such EEG "aftereffects" of transcorneal alternating current stimulation (tACS) at the chronic posttraumatic state in rats. EEG baseline was recorded from visual cortex under ketamine/xylazine narcosis of healthy rats and rats with chronic severe optic nerve crush. One week later, both groups were again anesthetized and stimulated transcorneally twice for 12 min each time. tACS-induced changes were compared with baseline EEG. Over the course of 65 min narcosis baseline EEG revealed a shift from a dominant delta power to theta. This shift was significantly delayed in lesioned animals compared with healthy controls. tACS applied during the late narcosis stage in normal rats led to significantly increased theta power with a parallel shift of the dominating peak to higher frequency which outlasted the stimulation period by 15 min (aftereffects). EEG in lesioned rats was not significantly changed. In rodents, tACS can induce neuroplasticity as shown by EEG aftereffects that outlast the stimulation period. But this requires a minimal level of brain activation because aftereffects are not seen when tACS is applied during deep anesthesia and not when applied to animals after severe optic nerve damage. We conclude that tACS is only effective to induce cortical plasticity when the the retina can be excited.