Sepsis is the third leading cause of death worldwide and the main cause of mortality in hospitals
, but the best treatment strategy remains uncertain. In particular, evidence suggests that current ...practices in the administration of intravenous fluids and vasopressors are suboptimal and likely induce harm in a proportion of patients
. To tackle this sequential decision-making problem, we developed a reinforcement learning agent, the Artificial Intelligence (AI) Clinician, which extracted implicit knowledge from an amount of patient data that exceeds by many-fold the life-time experience of human clinicians and learned optimal treatment by analyzing a myriad of (mostly suboptimal) treatment decisions. We demonstrate that the value of the AI Clinician's selected treatment is on average reliably higher than human clinicians. In a large validation cohort independent of the training data, mortality was lowest in patients for whom clinicians' actual doses matched the AI decisions. Our model provides individualized and clinically interpretable treatment decisions for sepsis that could improve patient outcomes.
To systematically examine the design, reporting standards, risk of bias, and claims of studies comparing the performance of diagnostic deep learning algorithms for medical imaging with that of expert ...clinicians.
Systematic review.
Medline, Embase, Cochrane Central Register of Controlled Trials, and the World Health Organization trial registry from 2010 to June 2019.
Randomised trial registrations and non-randomised studies comparing the performance of a deep learning algorithm in medical imaging with a contemporary group of one or more expert clinicians. Medical imaging has seen a growing interest in deep learning research. The main distinguishing feature of convolutional neural networks (CNNs) in deep learning is that when CNNs are fed with raw data, they develop their own representations needed for pattern recognition. The algorithm learns for itself the features of an image that are important for classification rather than being told by humans which features to use. The selected studies aimed to use medical imaging for predicting absolute risk of existing disease or classification into diagnostic groups (eg, disease or non-disease). For example, raw chest radiographs tagged with a label such as pneumothorax or no pneumothorax and the CNN learning which pixel patterns suggest pneumothorax.
Adherence to reporting standards was assessed by using CONSORT (consolidated standards of reporting trials) for randomised studies and TRIPOD (transparent reporting of a multivariable prediction model for individual prognosis or diagnosis) for non-randomised studies. Risk of bias was assessed by using the Cochrane risk of bias tool for randomised studies and PROBAST (prediction model risk of bias assessment tool) for non-randomised studies.
Only 10 records were found for deep learning randomised clinical trials, two of which have been published (with low risk of bias, except for lack of blinding, and high adherence to reporting standards) and eight are ongoing. Of 81 non-randomised clinical trials identified, only nine were prospective and just six were tested in a real world clinical setting. The median number of experts in the comparator group was only four (interquartile range 2-9). Full access to all datasets and code was severely limited (unavailable in 95% and 93% of studies, respectively). The overall risk of bias was high in 58 of 81 studies and adherence to reporting standards was suboptimal (<50% adherence for 12 of 29 TRIPOD items). 61 of 81 studies stated in their abstract that performance of artificial intelligence was at least comparable to (or better than) that of clinicians. Only 31 of 81 studies (38%) stated that further prospective studies or trials were required.
Few prospective deep learning studies and randomised trials exist in medical imaging. Most non-randomised trials are not prospective, are at high risk of bias, and deviate from existing reporting standards. Data and code availability are lacking in most studies, and human comparator groups are often small. Future studies should diminish risk of bias, enhance real world clinical relevance, improve reporting and transparency, and appropriately temper conclusions.
PROSPERO CRD42019123605.
Biofilms cause significant problems in the environment and during the treatment of infections. However, the molecular mechanisms underlying biofilm formation are poorly understood. There is a ...particular lack of knowledge about biofilm maturation processes, such as biofilm structuring and detachment, which are deemed crucial for the maintenance of biofilm viability and the dissemination of cells from a biofilm. Here, we identify the phenol-soluble modulin (PSM) surfactant peptides as key biofilm structuring factors in the premier biofilm-forming pathogen Staphylococcus aureus. We provide evidence that all known PSM classes participate in structuring and detachment processes. Specifically, absence of PSMs in isogenic S. aureus psm deletion mutants led to strongly impaired formation of biofilm channels, abolishment of the characteristic waves of biofilm detachment and regrowth, and loss of control of biofilm expansion. In contrast, induced expression of psm loci in preformed biofilms promoted those processes. Furthermore, PSMs facilitated dissemination from an infected catheter in a mouse model of biofilm-associated infection. Moreover, formation of the biofilm structure was linked to strongly variable, quorum sensing-controlled PSM expression in biofilm microenvironments, whereas overall PSM production remained constant to ascertain biofilm homeostasis. Our study describes a mechanism of biofilm structuring in molecular detail, and the general principle (i.e., quorum-sensing controlled expression of surfactants) seems to be conserved in several bacteria, despite the divergence of the respective biofilm-structuring surfactants. These findings provide a deeper understanding of biofilm development processes, which represents an important basis for strategies to interfere with biofilm formation in the environment and human disease.
Effective targeted therapy for sepsis requires an understanding of the heterogeneity in the individual host response to infection. We investigated this heterogeneity by defining interindividual ...variation in the transcriptome of patients with sepsis and related this to outcome and genetic diversity.
We assayed peripheral blood leucocyte global gene expression for a prospective discovery cohort of 265 adult patients admitted to UK intensive care units with sepsis due to community-acquired pneumonia and evidence of organ dysfunction. We then validated our findings in a replication cohort consisting of a further 106 patients. We mapped genomic determinants of variation in gene transcription between patients as expression quantitative trait loci (eQTL).
We discovered that following admission to intensive care, transcriptomic analysis of peripheral blood leucocytes defines two distinct sepsis response signatures (SRS1 and SRS2). The presence of SRS1 (detected in 108 41% patients in discovery cohort) identifies individuals with an immunosuppressed phenotype that included features of endotoxin tolerance, T-cell exhaustion, and downregulation of human leucocyte antigen (HLA) class II. SRS1 was associated with higher 14 day mortality than was SRS2 (discovery cohort hazard ratio (HR) 2·4, 95% CI 1·3–4·5, p=0·005; validation cohort HR 2·8, 95% CI 1·5–5·1, p=0·0007). We found that a predictive set of seven genes enabled the classification of patients as SRS1 or SRS2. We identified cis-acting and trans-acting eQTL for key immune and metabolic response genes and sepsis response networks. Sepsis eQTL were enriched in endotoxin-induced epigenetic marks and modulated the individual host response to sepsis, including effects specific to SRS group. We identified regulatory genetic variants involving key mediators of gene networks implicated in the hypoxic response and the switch to glycolysis that occurs in sepsis, including HIF1α and mTOR, and mediators of endotoxin tolerance, T-cell activation, and viral defence.
Our integrated genomics approach advances understanding of heterogeneity in sepsis by defining subgroups of patients with different immune response states and prognoses, as well as revealing the role of underlying genetic variation. Our findings provide new insights into the pathogenesis of sepsis and create opportunities for a precision medicine approach to enable targeted therapeutic intervention to improve sepsis outcomes.
European Commission, Medical Research Council (UK), and the Wellcome Trust.
Methicillin-resistant Staphylococcus aureus (MRSA) is a leading cause of morbidity and death. Phenol-soluble modulins (PSMs) are recently-discovered toxins with a key impact on the development of ...Staphylococcus aureus infections. Allelic variants of PSMs and their potential impact on pathogen success during infection have not yet been described. Here we show that the clonal complex (CC) 30 lineage, a major cause of hospital-associated sepsis and hematogenous complications, expresses an allelic variant of the PSMα3 peptide. We found that this variant, PSMα3N22Y, is characteristic of CC30 strains and has significantly reduced cytolytic and pro-inflammatory potential. Notably, CC30 strains showed reduced cytolytic and chemotactic potential toward human neutrophils, and increased hematogenous seeding in a bacteremia model, compared to strains in which the genome was altered to express non-CC30 PSMα3. Our findings describe a molecular mechanism contributing to attenuated pro-inflammatory potential in a main MRSA lineage. They suggest that reduced pathogen recognition via PSMs allows the bacteria to evade elimination by innate host defenses during bloodstream infections. Furthermore, they underscore the role of point mutations in key S. aureus toxin genes in that adaptation and the pivotal importance PSMs have in defining key S. aureus immune evasion and virulence mechanisms.
AbstractObjectiveTo systematically examine the design, reporting standards, risk of bias, and claims of studies comparing the performance of diagnostic deep learning algorithms for medical imaging ...with that of expert clinicians.DesignSystematic review.Data sourcesMedline, Embase, Cochrane Central Register of Controlled Trials, and the World Health Organization trial registry from 2010 to June 2019.Eligibility criteria for selecting studiesRandomised trial registrations and non-randomised studies comparing the performance of a deep learning algorithm in medical imaging with a contemporary group of one or more expert clinicians. Medical imaging has seen a growing interest in deep learning research. The main distinguishing feature of convolutional neural networks (CNNs) in deep learning is that when CNNs are fed with raw data, they develop their own representations needed for pattern recognition. The algorithm learns for itself the features of an image that are important for classification rather than being told by humans which features to use. The selected studies aimed to use medical imaging for predicting absolute risk of existing disease or classification into diagnostic groups (eg, disease or non-disease). For example, raw chest radiographs tagged with a label such as pneumothorax or no pneumothorax and the CNN learning which pixel patterns suggest pneumothorax.Review methodsAdherence to reporting standards was assessed by using CONSORT (consolidated standards of reporting trials) for randomised studies and TRIPOD (transparent reporting of a multivariable prediction model for individual prognosis or diagnosis) for non-randomised studies. Risk of bias was assessed by using the Cochrane risk of bias tool for randomised studies and PROBAST (prediction model risk of bias assessment tool) for non-randomised studies.ResultsOnly 10 records were found for deep learning randomised clinical trials, two of which have been published (with low risk of bias, except for lack of blinding, and high adherence to reporting standards) and eight are ongoing. Of 81 non-randomised clinical trials identified, only nine were prospective and just six were tested in a real world clinical setting. The median number of experts in the comparator group was only four (interquartile range 2-9). Full access to all datasets and code was severely limited (unavailable in 95% and 93% of studies, respectively). The overall risk of bias was high in 58 of 81 studies and adherence to reporting standards was suboptimal (<50% adherence for 12 of 29 TRIPOD items). 61 of 81 studies stated in their abstract that performance of artificial intelligence was at least comparable to (or better than) that of clinicians. Only 31 of 81 studies (38%) stated that further prospective studies or trials were required.ConclusionsFew prospective deep learning studies and randomised trials exist in medical imaging. Most non-randomised trials are not prospective, are at high risk of bias, and deviate from existing reporting standards. Data and code availability are lacking in most studies, and human comparator groups are often small. Future studies should diminish risk of bias, enhance real world clinical relevance, improve reporting and transparency, and appropriately temper conclusions.Study registrationPROSPERO CRD42019123605.
Effective therapies for patients with coronavirus disease 2019 (COVID-19) are needed, and clinical trial data have demonstrated that low-dose dexamethasone reduced mortality in hospitalized patients ...with COVID-19 who required respiratory support.
To estimate the association between administration of corticosteroids compared with usual care or placebo and 28-day all-cause mortality.
Prospective meta-analysis that pooled data from 7 randomized clinical trials that evaluated the efficacy of corticosteroids in 1703 critically ill patients with COVID-19. The trials were conducted in 12 countries from February 26, 2020, to June 9, 2020, and the date of final follow-up was July 6, 2020. Pooled data were aggregated from the individual trials, overall, and in predefined subgroups. Risk of bias was assessed using the Cochrane Risk of Bias Assessment Tool. Inconsistency among trial results was assessed using the I2 statistic. The primary analysis was an inverse variance-weighted fixed-effect meta-analysis of overall mortality, with the association between the intervention and mortality quantified using odds ratios (ORs). Random-effects meta-analyses also were conducted (with the Paule-Mandel estimate of heterogeneity and the Hartung-Knapp adjustment) and an inverse variance-weighted fixed-effect analysis using risk ratios.
Patients had been randomized to receive systemic dexamethasone, hydrocortisone, or methylprednisolone (678 patients) or to receive usual care or placebo (1025 patients).
The primary outcome measure was all-cause mortality at 28 days after randomization. A secondary outcome was investigator-defined serious adverse events.
A total of 1703 patients (median age, 60 years interquartile range, 52-68 years; 488 29% women) were included in the analysis. Risk of bias was assessed as "low" for 6 of the 7 mortality results and as "some concerns" in 1 trial because of the randomization method. Five trials reported mortality at 28 days, 1 trial at 21 days, and 1 trial at 30 days. There were 222 deaths among the 678 patients randomized to corticosteroids and 425 deaths among the 1025 patients randomized to usual care or placebo (summary OR, 0.66 95% CI, 0.53-0.82; P < .001 based on a fixed-effect meta-analysis). There was little inconsistency between the trial results (I2 = 15.6%; P = .31 for heterogeneity) and the summary OR was 0.70 (95% CI, 0.48-1.01; P = .053) based on the random-effects meta-analysis. The fixed-effect summary OR for the association with mortality was 0.64 (95% CI, 0.50-0.82; P < .001) for dexamethasone compared with usual care or placebo (3 trials, 1282 patients, and 527 deaths), the OR was 0.69 (95% CI, 0.43-1.12; P = .13) for hydrocortisone (3 trials, 374 patients, and 94 deaths), and the OR was 0.91 (95% CI, 0.29-2.87; P = .87) for methylprednisolone (1 trial, 47 patients, and 26 deaths). Among the 6 trials that reported serious adverse events, 64 events occurred among 354 patients randomized to corticosteroids and 80 events occurred among 342 patients randomized to usual care or placebo.
In this prospective meta-analysis of clinical trials of critically ill patients with COVID-19, administration of systemic corticosteroids, compared with usual care or placebo, was associated with lower 28-day all-cause mortality.
Despite progress in the supportive care available for critically ill patients, few advances have been made in the search for effective disease-modifying therapeutic options. The fact that many trials ...in critical care medicine have not identified a treatment benefit is probably due, in part, to the underlying heterogeneity of critical care syndromes. Numerous approaches have been proposed to divide populations of critically ill patients into more meaningful subgroups (subphenotypes), some of which might be more useful than others. Subclassification systems driven by clinical features and biomarkers have been proposed for acute respiratory distress syndrome, sepsis, acute kidney injury, and pancreatitis. Identifying the systems that are most useful and biologically meaningful could lead to a better understanding of the pathophysiology of critical care syndromes and the discovery of new treatment targets, and allow recruitment in future therapeutic trials to focus on predicted responders. This Review discusses proposed subphenotypes of critical illness syndromes and highlights the issues that will need to be addressed to translate subphenotypes into clinical practice.
PDF
