SynthEye: Investigating the Impact of Synthetic Data on Artificial Intelligence-assisted Gene Diagnosis of Inherited Retinal Disease

Purpose

Rare disease diagnosis is challenging in medical image-based artificial intelligence due to a natural class imbalance in datasets, leading to biased prediction models. Inherited retinal diseases (IRDs) are a research domain that particularly faces this issue. This study investigates the applicability of synthetic data in improving artificial intelligence-enabled diagnosis of IRDs using generative adversarial networks (GANs).

Design

Diagnostic study of gene-labeled fundus autofluorescence (FAF) IRD images using deep learning.

Participants

Moorfields Eye Hospital (MEH) dataset of 15 692 FAF images obtained from 1800 patients with confirmed genetic diagnosis of 1 of 36 IRD genes.

Methods

A StyleGAN2 model is trained on the IRD dataset to generate 512 × 512 resolution images. Convolutional neural networks are trained for classification using different synthetically augmented datasets, including real IRD images plus 1800 and 3600 synthetic images, and a fully rebalanced dataset. We also perform an experiment with only synthetic data. All models are compared against a baseline convolutional neural network trained only on real data.

Main Outcome Measures

We evaluated synthetic data quality using a Visual Turing Test conducted with 4 ophthalmologists from MEH. Synthetic and real images were compared using feature space visualization, similarity analysis to detect memorized images, and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) score for no-reference-based quality evaluation. Convolutional neural network diagnostic performance was determined on a held-out test set using the area under the receiver operating characteristic curve (AUROC) and Cohen's Kappa (κ).

Results

An average true recognition rate of 63% and fake recognition rate of 47% was obtained from the Visual Turing Test. Thus, a considerable proportion of the synthetic images were classified as real by clinical experts. Similarity analysis showed that the synthetic images were not copies of the real images, indicating that copied real images, meaning the GAN was able to generalize. However, BRISQUE score analysis indicated that synthetic images were of significantly lower quality overall than real images (P < 0.05). Comparing the rebalanced model (RB) with the baseline (R), no significant change in the average AUROC and κ was found (R-AUROC = 0.86[0.85-88], RB-AUROC = 0.88[0.86-0.89], R-k = 0.51[0.49-0.53], and RB-k = 0.52[0.50-0.54]). The synthetic data trained model (S) achieved similar performance as the baseline (S-AUROC = 0.86[0.85-87], S-k = 0.48[0.46-0.50]).

Conclusions

Synthetic generation of realistic IRD FAF images is feasible. Synthetic data augmentation does not deliver improvements in classification performance. However, synthetic data alone deliver a similar performance as real data, and hence may be useful as a proxy to real data. Financial Disclosure(s): Proprietary or commercial disclosure may be found after the references.

Impact of ammonia levels on outcome in clinically stable outpatients with advanced chronic liver disease

Background & Aims

Ammonia levels predicted hospitalisation in a recent landmark study not accounting for portal hypertension and systemic inflammation severity. We investigated (i) the prognostic value of venous ammonia levels (outcome cohort) for liver-related outcomes while accounting for these factors and (ii) its correlation with key disease-driving mechanisms (biomarker cohort).

Methods

(i) The outcome cohort included 549 clinically stable outpatients with evidence of advanced chronic liver disease. (ii) The partly overlapping biomarker cohort comprised 193 individuals, recruited from the prospective Vienna Cirrhosis Study (VICIS: NCT03267615).

Results

(i) In the outcome cohort, ammonia increased across clinical stages as well as hepatic venous pressure gradient and United Network for Organ Sharing model for end-stage liver disease (2016) strata and were independently linked with diabetes. Ammonia was associated with liver-related death, even after multivariable adjustment (adjusted hazard ratio [aHR]: 1.05 [95% CI: 1.00-1.10]; p = 0.044). The recently proposed cut-off (≥1.4 × upper limit of normal) was independently predictive of hepatic decompensation (aHR: 2.08 [95% CI: 1.35-3.22]; p <0.001), non-elective liver-related hospitalisation (aHR: 1.86 [95% CI: 1.17-2.95]; p = 0.008), and - in those with decompensated advanced chronic liver disease - acute-on-chronic liver failure (aHR: 1.71 [95% CI: 1.05-2.80]; p = 0.031). (ii) Besides hepatic venous pressure gradient, venous ammonia was correlated with markers of endothelial dysfunction and liver fibrogenesis/matrix remodelling in the biomarker cohort.

Conclusions

Venous ammonia predicts hepatic decompensation, non-elective liver-related hospitalisation, acute-on-chronic liver failure, and liver-related death, independently of established prognostic indicators including C-reactive protein and hepatic venous pressure gradient. Although venous ammonia is linked with several key disease-driving mechanisms, its prognostic value is not explained by associated hepatic dysfunction, systemic inflammation, or portal hypertension severity, suggesting direct toxicity.

Impact and implications

A recent landmark study linked ammonia levels (a simple blood test) with hospitalisation/death in individuals with clinically stable cirrhosis. Our study extends the prognostic value of venous ammonia to other important liver-related complications. Although venous ammonia is linked with several key disease-driving mechanisms, they do not fully explain its prognostic value. This supports the concept of direct ammonia toxicity and ammonia-lowering drugs as disease-modifying treatment.

The NAFLD Decompensation Risk Score: External Validation and Comparison to Existing Models to Predict Hepatic Events in a Retrospective Cohort Study

Background

The NAFLD decompensation risk score (the Iowa Model) was recently developed to identify patients with nonalcoholic fatty liver disease (NAFLD) at highest risk of developing hepatic events using three variables-age, platelet count, and diabetes.

Aims

We performed an external validation of the Iowa Model and compared it to existing non-invasive models.

Methods

We included 249 patients with NAFLD at Boston Medical Center, Boston, Massachusetts, in the external validation cohort and 949 patients in the combined internal/external validation cohort. The primary outcome was the development of hepatic events (ascites, hepatic encephalopathy, esophageal or gastric varices, or hepatocellular carcinoma). We used Cox proportional hazards to analyze the ability of the Iowa Model to predict hepatic events in the external validation (https://uihc.org/non-alcoholic-fatty-liver-disease-decompensation-risk-score-calculator). We compared the performance of the Iowa Model to the AST-to-platelet ratio index (APRI), NAFLD fibrosis score (NFS), and the FIB-4 index in the combined cohort.

Results

The Iowa Model significantly predicted the development of hepatic events with hazard ratio of 2.5 [95% confidence interval (CI) 1.7-3.9, P < 0.001] and area under the receiver operating characteristic curve (AUROC) of 0.87 (CI 0.83-0.91). The AUROC of the Iowa Model (0.88, CI: 0.85-0.92) was comparable to the FIB-4 index (0.87, CI: 0.83-0.91) and higher than NFS (0.66, CI: 0.63-0.69) and APRI (0.76, CI: 0.73-0.79).

Conclusions

In an urban, racially and ethnically diverse population, the Iowa Model performed well to identify NAFLD patients at higher risk for liver-related complications. The model provides the individual probability of developing hepatic events and identifies patients in need of early intervention.

Stellate cell expression of SPARC-related modular calcium-binding protein 2 is associated with human non-alcoholic fatty liver disease severity

Background & Aims

Histological assessment of liver biopsies is the gold standard for diagnosis of non-alcoholic steatohepatitis (NASH), the progressive form of non-alcoholic fatty liver disease (NAFLD), despite its well-established limitations. Therefore, non-invasive biomarkers that can offer an integrated view of the liver are needed to improve diagnosis and reduce sampling bias. Hepatic stellate cells (HSCs) are central in the development of hepatic fibrosis, a hallmark of NASH. Secreted HSC-specific proteins may, therefore, reflect disease state in the NASH liver and serve as non-invasive diagnostic biomarkers.

Methods

We performed RNA-sequencing on liver biopsies from a histologically characterised cohort of obese patients (n = 30, BMI >35 kg/m²) to identify and evaluate HSC-specific genes encoding secreted proteins. Bioinformatics was used to identify potential biomarkers and their expression at single-cell resolution. We validated our findings using single-molecule fluorescence in situ hybridisation (smFISH) and ELISA to detect mRNA in liver tissue and protein levels in plasma, respectively.

Results

Hepatic expression of SPARC-related modular calcium-binding protein 2 (SMOC2) was increased in NASH compared to no-NAFLD (p.adj <0.001). Single-cell RNA-sequencing data indicated that SMOC2 was primarily expressed by HSCs, which was validated using smFISH. Finally, plasma SMOC2 was elevated in NASH compared to no-NAFLD (p <0.001), with a predictive accuracy of AUROC 0.88.

Conclusions

Increased SMOC2 in plasma appears to reflect HSC activation, a key cellular event associated with NASH progression, and may serve as a non-invasive biomarker of NASH.

Impact and implications

Non-alcoholic fatty liver disease (NAFLD) and its progressive form, non-alcoholic steatohepatitis (NASH), are the most common forms of chronic liver diseases. Currently, liver biopsies are the gold standard for diagnosing NAFLD. Blood-based biomarkers to complement liver biopsies for diagnosis of NAFLD are required. We found that activated hepatic stellate cells, a cell type central to NAFLD pathogenesis, upregulate expression of the secreted protein SPARC-related modular calcium-binding protein 2 (SMOC2). SMOC2 was elevated in blood samples from patients with NASH and may hold promise as a blood-based biomarker for the diagnosis of NAFLD.

Predicting the survival benefit of liver transplantation in HBV-related acute-on-chronic liver failure: an observational cohort study

Background

Liver transplantation (LT) is an effective therapy for acute-on-chronic liver failure (ACLF) but is limited by organ shortages. We aimed to identify an appropriate score for predicting the survival benefit of LT in HBV-related ACLF patients.

Methods

Hospitalized patients with acute deterioration of HBV-related chronic liver disease (n = 4577) from the Chinese Group on the Study of Severe Hepatitis B (COSSH) open cohort were enrolled to evaluate the performance of five commonly used scores for predicting the prognosis and transplant survival benefit. The survival benefit rate was calculated to reflect the extended rate of the expected lifetime with vs. without LT.

Findings

In total, 368 HBV-ACLF patients received LT. They showed significantly higher 1-year survival than those on the waitlist in both the entire HBV-ACLF cohort (77.2%/52.3%, p < 0.001) and the propensity score matching cohort (77.2%/27.6%, p < 0.001). The area under the receiver operating characteristic curve (AUROC) showed that the COSSH-ACLF II score performed best (AUROC 0.849) at identifying the 1-year risk of death on the waitlist and best (AUROC 0.864) at predicting 1-year outcome post-LT (COSSH-ACLFs/CLIF-C ACLFs/MELDs/MELD-Nas: AUROC 0.835/0.825/0.796/0.781; all p < 0.05). The C-indexes confirmed the high predictive value of COSSH-ACLF IIs. Survival benefit rate analyses showed that patients with COSSH-ACLF IIs 7-10 had a higher 1-year survival benefit rate from LT (39.2%-64.3%) than those with score <7 or >10. These results were prospectively validated.

Interpretation

COSSH-ACLF IIs identified the risk of death on the waitlist and accurately predicted post-LT mortality and survival benefit for HBV-ACLF. Patients with COSSH-ACLF IIs 7-10 derived a higher net survival benefit from LT.

Funding

This study was supported by the National Natural Science Foundation of China (No. 81830073, No. 81771196) and the National Special Support Program for High-Level Personnel Recruitment (Ten-thousand Talents Program).

Real-Time Prediction of Mortality, Cardiac Arrest, and Thromboembolic Complications in Hospitalized Patients With COVID-19

Background

COVID-19 infection carries significant morbidity and mortality. Current risk prediction for complications in COVID-19 is limited, and existing approaches fail to account for the dynamic course of the disease.

Objectives

The purpose of this study was to develop and validate the COVID-HEART predictor, a novel continuously updating risk-prediction technology to forecast adverse events in hospitalized patients with COVID-19.

Methods

Retrospective registry data from patients with severe acute respiratory syndrome coronavirus 2 infection admitted to 5 hospitals were used to train COVID-HEART to predict all-cause mortality/cardiac arrest (AM/CA) and imaging-confirmed thromboembolic events (TEs) (n = 2,550 and n = 1,854, respectively). To assess COVID-HEART's performance in the face of rapidly changing clinical treatment guidelines, an additional 1,100 and 796 patients, admitted after the completion of development data collection, were used for testing. Leave-hospital-out validation was performed.

Results

Over 20 iterations of temporally divided testing, the mean area under the receiver operating characteristic curve were 0.917 (95% confidence interval [CI]: 0.916-0.919) and 0.757 (95% CI: 0.751-0.763) for prediction of AM/CA and TE, respectively. The interquartile ranges of median early warning times were 14 to 21 hours for AM/CA and 12 to 60 hours for TE. The mean area under the receiver operating characteristic curve for the left-out hospitals were 0.956 (95% CI: 0.936-0.976) and 0.781 (95% CI: 0.642-0.919) for prediction of AM/CA and TE, respectively.

Conclusions

The continuously updating, fully interpretable COVID-HEART predictor accurately predicts AM/CA and TE within multiple time windows in hospitalized COVID-19 patients. In its current implementation, the predictor can facilitate practical, meaningful changes in patient triage and resource allocation by providing real-time risk scores for these outcomes. The potential utility of the predictor extends to COVID-19 patients after hospitalization and beyond COVID-19.

Prospective comparison of transient elastography, MRI and serum scores for grading steatosis and detecting non-alcoholic steatohepatitis in bariatric surgery candidates

Background & Aims

Tools for the non-invasive diagnosis of non-alcoholic steatohepatitis (NASH) in morbidly obese patients with suspected non-alcoholic fatty liver disease (NAFLD) are an unmet clinical need. We prospectively compared the performance of transient elastography, MRI, and 3 serum scores for the diagnosis of NAFLD, grading of steatosis and detection of NASH in bariatric surgery candidates.

Methods

Of 186 patients screened, 152 underwent liver biopsy, which was used as a reference for NAFLD (steatosis [S]>5%), steatosis grading and NASH diagnosis. Biopsies were read by a single expert pathologist. MRI-based proton density fat fraction (MRI-PDFF) was measured in an open-bore, vertical field 1.0T scanner and controlled attenuation parameter (CAP) was measured by transient elastography, using the XL probe. Serum scores (SteatoTest, hepatic steatosis index and fatty liver index) were also calculated.

Results

The applicability of MRI was better than that of FibroScan (98% vs. 79%; p <0.0001). CAP had AUROCs of 0.83, 0.79, 0.73 and 0.69 for S>5%, S>33%, S>66% and NASH, respectively. Transient elastography had an AUROC of 0.80 for significant fibrosis (F0-F1 vs. F2-F3). MRI-PDFF had AUROCs of 0.97, 0.95, 0.92 and 0.84 for S>5%, S>33%, S>66% and NASH, respectively. When compared head-to-head in the 97 patients with all valid tests available, MRI-PDFF outperformed CAP for grading steatosis (S>33%, AUROC 0.97 vs. 0.78; p <0.0003 and S>66%, AUROC 0.93 vs. 0.75; p = 0.0015) and diagnosing NASH (AUROC 0.82 vs. 0.68; p = 0.0056). When compared in "intention to diagnose" analysis, MRI-PDFF outperformed CAP, hepatic steatosis index and fatty liver index for grading steatosis (S>5%, S>33% and S>66%).

Conclusion

MRI-PDFF outperforms CAP for diagnosing NAFLD, grading steatosis and excluding NASH in morbidly obese patients undergoing bariatric surgery.

Lay summary

Non-invasive tests for detecting fatty liver and steatohepatitis, the active form of the disease, have not been well studied in obese patients who are candidates for bariatric surgery. The most popular tests for this purpose are Fibroscan, which can be used to measure the controlled attenuation parameter (CAP), and magnetic resonance imaging, which can be used to measure the proton density fat fraction (MRI-PDFF). We found that, when taking liver biopsy as a reference, MRI-PDFF performed better than CAP for detecting and grading fatty liver as well as excluding steatohepatitis in morbidly obese patients undergoing bariatric surgery.

Admission Serum Urea is a Better Predictor of Mortality than Creatinine in Patients With Acute-On-Chronic Liver Failure and Acute Kidney Injury

BACKGROUND

The occurrence of acute kidney injury (AKI) in acute-on-chronic liver failure (ACLF) negatively impacts the survival of patients. There are scant data on the impact of serum urea on outcomes in these patients. We performed this study to evaluate the relationship between admission serum urea and the survival in patients with ACLF and AKI.

METHODS

A prospective study was conducted on patients with ACLF (as per Asian Pacific Association for the Study of the Liver criteria) and AKI (as per Acute Kidney Injury Network criteria) hospitalized in the gastroenterology ward between October 2016 and May 2018. Demographic, clinical and laboratory parameters were recorded, and outcomes were compared in patients with respect to the admission serum urea level.

RESULTS

A total of 103 of 143 hospitalized patients with ACLF had AKI and were included as study subjects. The discrimination ability between survivors and the deceased was similar for serum urea levels (area under the receiver operating characteristic curve [AUROC] [95% confidence interval {CI}]: 28 days survival, 0.76 [0.67-0.85]; 90 days survival, 0.81 [0.72-0.91]) and serum creatinine levels (AUROC [95% CI]: 28 days survival, 0.75 [0.66-0.84]; 90 days survival: 0.77 [0.67-0.88]) in patients with ACLF and AKI. However, on multivariate analysis, admission serum urea (not serum creatinine) was an independent predictor of mortality in these patients both at 28 days (p = 0.001, adjusted hazard ratio [AHR]: 1.013 [1.005-1.021]) and 90 days (p = 0.001, AHR: 1.014 [1.006-1.022]).

CONCLUSION

Over two-thirds of patients with ACLF had AKI. The discrimination ability between survivors and the deceased was similar for both serum urea and serum creatinine levels. However admission serum urea was found to be a better predictor of mortality than serum creatinine in patients with ACLF and AKI.

Validation of a Retrospective Computing Model for Mortality Risk in the Intensive Care Unit

Objective

To compare the predictive performance of Epic Systems Corporation’s proprietary intensive care unit (ICU) mortality risk model (IMRM) with that of the Acute Physiology and Chronic Health Evaluation (APACHE) IV score.

Methods

This is a retrospective cohort study of patients treated from January 1, 2008, through January 1, 2018. This single-center study was performed at Mayo Clinic (Rochester, MN), a tertiary care teaching and referral center. The primary outcome was death in the ICU. Discrimination of each risk model for hospital mortality was assessed by comparing area under the receiver operating characteristic curve (AUROC).

Results

The cohort mostly comprised older patients (median age, 64 years) and men (56.7%). The mortality rate of the cohort was 3.5% (2251 of 63,775 patients). The AUROC for mortality prediction was 89.7% (95% CI, 89.5% to 89.9%) for the IMRM, which was significantly greater than the AUROC of 88.2% (95% CI, 87.9% to 88.4%) for APACHE IV (P<.001).

Conclusion

The IMRM was superior to the commonly used APACHE IV score and may be easily integrated into electronic health records at any hospital using Epic software.

The liver matrisome - looking beyond collagens

The extracellular matrix (ECM) is a diverse microenvironment that maintains bidirectional communication with surrounding cells to regulate cell and tissue homeostasis. The classical definition of the ECM has more recently been extended to include non-fibrillar proteins that either interact or are structurally affiliated with the ECM, termed the 'matrisome.' In addition to providing the structure and architectural support for cells and tissue, the matrisome serves as a reservoir for growth factors and cytokines, as well as a signaling hub via which cells can communicate with their environment and vice-versa. The matrisome is a master regulator of tissue homeostasis and organ function, which can dynamically and appropriately respond to any stress or injury. Failure to properly regulate these responses can lead to changes in the matrisome that are maladaptive. Hepatic fibrosis is a canonical example of ECM dyshomeostasis, leading to accumulation of predominantly collagenous ECM; indeed, hepatic fibrosis is considered almost synonymous with collagen accumulation. However, the qualitative and quantitative alterations of the hepatic matrisome during fibrosis are much more diverse than simple accumulation of collagens and occur long before fibrosis is histologically detected. A deeper understanding of the hepatic matrisome and its response to injury could yield new mechanistic insights into disease progression and regression, as well as potentially identify new biomarkers for both. In this review, we discuss the role of the ECM in liver diseases and look at new "omic" approaches to study this compartment.

Natriuretic Peptides to Predict Short-Term Mortality in Patients With Sepsis: A Systematic Review and Meta-analysis

Data are conflicting regarding the optimal cutoffs of B-type natriuretic peptide (BNP) and N-terminal pro-B-type natriuretic peptide (NT-proBNP) to predict short-term mortality in patients with sepsis. We conducted a comprehensive search of several databases (MEDLINE, EMBASE, Cochrane Central Register of Controlled Trials, Cochrane Database of Systematic Reviews, and Scopus) for English-language reports of studies evaluating adult patients with sepsis, severe sepsis, and septic shock with BNP/NT-proBNP levels and short-term mortality (intensive care unit, in-hospital, 28-day, or 30-day) published from January 1, 2000, to September 5, 2017. The average values in survivors and nonsurvivors were used to estimate the receiver operating characteristic curve (ROC) using a parametric regression model. Thirty-five observational studies (3508 patients) were included (median age, 51-75 years; 12%-74% males; cumulative mortality, 34.2%). A BNP of 622 pg/mL had the greatest discrimination for mortality (sensitivity, 0.695 [95% CI, 0.659-0.729]; specificity, 0.907 [95% CI, 0.810-1.003]; area under the ROC, 0.766 [95% CI, 0.734-0.797]). An NT-proBNP of 4000 pg/mL had the greatest discrimination for mortality (sensitivity, 0.728 [95% CI, 0.703-0.753]; specificity, 0.789 [95% CI, 0.710-0.867]; area under the ROC, 0.787 [95% CI, 0.766-0.809]). In prespecified subgroup analyses, identified BNP/NT-proBNP cutoffs had higher discrimination if specimens were obtained 24 hours or less after admission, in patients with severe sepsis/septic shock, in patients enrolled after 2010, and in studies performed in the United States and Europe. There was inconsistent adjustment for renal function. In this hypothesis-generating analysis, BNP and NT-proBNP cutoffs of 622 pg/mL and 4000 pg/mL optimally predicted short-term mortality in patients with sepsis. The applicability of these results is limited by the heterogeneity of included patient populations.

Visual assessment of biliary excretion of Gd-EOB-DTPA in patients with suspected diffuse liver disease - A biopsy-verified prospective study

OBJECTIVES

To qualitatively evaluate late dynamic contrast phases, 10, 20 and 30 min, after administration of Gd-EOB-DTPA with regard to biliary excretion in patients presenting with elevated liver enzymes without clinical signs of cirrhosis or hepatic decompensation and to compare the visual assessment of contrast agent excretion with histo-pathological fibrosis stage, contrast uptake parameters and blood tests.

METHODS

29 patients were prospectively examined using 1.5 T MRI. The visually assessed presence or absence of contrast agent for each of five anatomical regions in randomly reviewed time-series was summarized on a four grade scale for each patient. The scores, including a total visual score, were related to the histo-pathological findings, the quantitative contrast agent uptake parameters, expressed as K Hep or LSC_N, and blood tests.

RESULTS

No relationship between the fibrosis grade or contrast uptake parameters could be established. A negative correlation between the visual assessment and alkaline phosphatase (ALP) was found. Comparing a sub-group of cholestatic patients with fibrosis score and Gd-EOB-DTPA dynamic parameters did not add any additional significant correlation.

CONCLUSIONS

No correlation between visually assessed biliary excretion of Gd-EOB-DTPA and histo-pathological or contrast uptake parameters was found. A negative correlation between the visual assessment and alkaline phosphatase (ALP) was found.