Deep learning generates synthetic cancer histology for explainability and education.

Artificial intelligence methods including deep neural networks (DNN) can provide rapid molecular classification of tumors from routine histology with accuracy that matches or exceeds human pathologists. Discerning how neural networks make their predictions remains a significant challenge, but explainability tools help provide insights into what models have learned when corresponding histologic features are poorly defined. Here, we present a method for improving explainability of DNN models using synthetic histology generated by a conditional generative adversarial network (cGAN). We show that cGANs generate high-quality synthetic histology images that can be leveraged for explaining DNN models trained to classify molecularly-subtyped tumors, exposing histologic features associated with molecular state. Fine-tuning synthetic histology through class and layer blending illustrates nuanced morphologic differences between tumor subtypes. Finally, we demonstrate the use of synthetic histology for augmenting pathologist-in-training education, showing that these intuitive visualizations can reinforce and improve understanding of histologic manifestations of tumor biology.

Association of Insurance Type With Inpatient Surgery 30-Day Complications and Costs.

INTRODUCTION: Safety-net hospitals (SNHs) have higher postoperative complications and costs versus low-burden hospitals. Do low socioeconomic status/vulnerable patients receive care at lower-quality hospitals or are there factors beyond providers' control? We studied the association of private, Medicare, and vulnerable insurance type with complications/costs in a high-burden SNH. METHODS: Retrospective inpatient cohort study using National Surgical Quality Improvement Program (NSQIP) data (2013-2019) with cost data risk-adjusted by frailty, preoperative serious acute conditions (PASC), case status, and expanded operative stress score (OSS) to evaluate 30-day unplanned reoperations, any complication, Clavien-Dindo IV (CDIV) complications, and hospitalization variable costs. RESULTS: Cases (Private 1517; Medicare 1224; Vulnerable 3648) with patient mean age 52.3 y [standard deviation = 14.7] and 47.3% male. Adjusting for frailty and OSS, vulnerable patients had higher odds of PASC (aOR = 1.71, CI = 1.39-2.10, P < 0.001) versus private. Adjusting for frailty, PASC and OSS, Medicare (aOR = 1.27, CI = 1.06-1.53, P = 0.009), and vulnerable (aOR = 2.44, CI = 2.13-2.79, P < 0.001) patients were more likely to undergo urgent/emergent surgeries. Vulnerable patients had increased odds of reoperation and any complications versus private. Variable cost percentage change was similar between private and vulnerable after adjusting for case status. Urgent/emergent case status increased percentage change costs by 32.31%. We simulated "switching" numbers of private (3648) versus vulnerable (1517) cases resulting in an estimated variable cost of $49.275 million, a 25.2% decrease from the original $65.859 million. CONCLUSIONS: Increased presentation acuity (PASC and urgent/emergent surgeries) in vulnerable patients drive increased odds of complications and costs versus private, suggesting factors beyond providers' control. The greatest impact on outcomes may be from decreasing the incidence of urgent/emergent surgeries by improving access to care.

Association of Socioeconomic Area Deprivation Index with Hospital Readmissions After Colon and Rectal Surgery.

BACKGROUND: Risk adjustment for reimbursement and quality measures omits social risk factors despite adversely affecting health outcomes. Social risk factors are not usually available in electronic health records (EHR) or administrative data. Socioeconomic status can be assessed by using US Census data. Distressed Communities Index (DCI) is based upon zip codes, and the Area Deprivation Index (ADI) provides more granular estimates at the block group level. We examined the association of neighborhood disadvantage using the ADI, DCI, and patient-level insurance status on 30-day readmission risk after colorectal surgery. METHODS: Our 677 patient cohort was derived from the 2013-2017 National Surgical Quality Improvement Program at a safety net hospital augmented with EHR data to determine insurance status and 30-day readmissions. Patients' home addresses were linked to the ADI and DCI. RESULTS: Our cohort consisted of 53.9% males and 63.8% Hispanics with a 22.9% 30-day readmission rate from the date of discharge; > 50% lived in highly deprived neighborhoods. Controlling for medical comorbidities and complications, ADI was associated with increased risk of 30 days from the date of discharge readmissions among patients living in medium (OR = 2.15, p = .02) or high (OR = 1.88, p = .03) deprived areas compared to less-deprived neighborhoods, but not insurance status or DCI. CONCLUSIONS: The ADI identified patients living in deprived communities with increased readmission risk. Our results show that block-group level ADI can potentially be used in risk adjustment, to identify high-risk patients and to design better care pathways that improve health outcomes.