Improving explainable AI with patch perturbation-based evaluation pipeline: a COVID-19 X-ray image analysis case study.

Recent advances in artificial intelligence (AI) have sparked interest in developing explainable AI (XAI) methods for clinical decision support systems, especially in translational research. Although using XAI methods may enhance trust in black-box models, evaluating their effectiveness has been challenging, primarily due to the absence of human (expert) intervention, additional annotations, and automated strategies. In order to conduct a thorough assessment, we propose a patch perturbation-based approach to automatically evaluate the quality of explanations in medical imaging analysis. To eliminate the need for human efforts in conventional evaluation methods, our approach executes poisoning attacks during model retraining by generating both static and dynamic triggers. We then propose a comprehensive set of evaluation metrics during the model inference stage to facilitate the evaluation from multiple perspectives, covering a wide range of correctness, completeness, consistency, and complexity. In addition, we include an extensive case study to showcase the proposed evaluation strategy by applying widely-used XAI methods on COVID-19 X-ray imaging classification tasks, as well as a thorough review of existing XAI methods in medical imaging analysis with evaluation availability. The proposed patch perturbation-based workflow offers model developers an automated and generalizable evaluation strategy to identify potential pitfalls and optimize their proposed explainable solutions, while also aiding end-users in comparing and selecting appropriate XAI methods that meet specific clinical needs in real-world clinical research and practice.

Early and fair COVID-19 outcome risk assessment using robust feature selection.

Personalized medicine plays an important role in treatment optimization for COVID-19 patient management. Early treatment in patients at high risk of severe complications is vital to prevent death and ventilator use. Predicting COVID-19 clinical outcomes using machine learning may provide a fast and data-driven solution for optimizing patient care by estimating the need for early treatment. In addition, it is essential to accurately predict risk across demographic groups, particularly those underrepresented in existing models. Unfortunately, there is a lack of studies demonstrating the equitable performance of machine learning models across patient demographics. To overcome this existing limitation, we generate a robust machine learning model to predict patient-specific risk of death or ventilator use in COVID-19 positive patients using features available at the time of diagnosis. We establish the value of our solution across patient demographics, including gender and race. In addition, we improve clinical trust in our automated predictions by generating interpretable patient clustering, patient-level clinical feature importance, and global clinical feature importance within our large real-world COVID-19 positive patient dataset. We achieved 89.38% area under receiver operating curve (AUROC) performance for severe outcomes prediction and our robust feature ranking approach identified the presence of dementia as a key indicator for worse patient outcomes. We also demonstrated that our deep-learning clustering approach outperforms traditional clustering in separating patients by severity of outcome based on mutual information performance. Finally, we developed an application for automated and fair patient risk assessment with minimal manual data entry using existing data exchange standards.

Explainable synthetic image generation to improve risk assessment of rare pediatric heart transplant rejection.

Expert microscopic analysis of cells obtained from frequent heart biopsies is vital for early detection of pediatric heart transplant rejection to prevent heart failure. Detection of this rare condition is prone to low levels of expert agreement due to the difficulty of identifying subtle rejection signs within biopsy samples. The rarity of pediatric heart transplant rejection also means that very few gold-standard images are available for developing machine learning models. To solve this urgent clinical challenge, we developed a deep learning model to automatically quantify rejection risk within digital images of biopsied tissue using an explainable synthetic data augmentation approach. We developed this explainable AI framework to illustrate how our progressive and inspirational generative adversarial network models distinguish between normal tissue images and those containing cellular rejection signs. To quantify biopsy-level rejection risk, we first detect local rejection features using a binary image classifier trained with expert-annotated and synthetic examples. We converted these local predictions into a biopsy-wide rejection score via an interpretable histogram-based approach. Our model significantly improves upon prior works with the same dataset with an area under the receiver operating curve (AUROC) of 98.84% for the local rejection detection task and 95.56% for the biopsy-rejection prediction task. A biopsy-level sensitivity of 83.33% makes our approach suitable for early screening of biopsies to prioritize expert analysis. Our framework provides a solution to rare medical imaging challenges currently limited by small datasets.

Identification of the Transcription Factor Relationships Associated with Androgen Deprivation Therapy Response and Metastatic Progression in Prostate Cancer.

BACKGROUND: Patients with locally advanced or recurrent prostate cancer typically undergo androgen deprivation therapy (ADT), but the benefits are often short-lived and the responses variable. ADT failure results in castration-resistant prostate cancer (CRPC), which inevitably leads to metastasis. We hypothesized that differences in tumor transcriptional programs may reflect differential responses to ADT and subsequent metastasis. RESULTS: We performed whole transcriptome analysis of 20 patient-matched Pre-ADT biopsies and 20 Post-ADT prostatectomy specimens, and identified two subgroups of patients (high impact and low impact groups) that exhibited distinct transcriptional changes in response to ADT. We found that all patients lost the AR-dependent subtype (PCS2) transcriptional signatures. The high impact group maintained the more aggressive subtype (PCS1) signal, while the low impact group more resembled an AR-suppressed (PCS3) subtype. Computational analyses identified transcription factor coordinated groups (TFCGs) enriched in the high impact group network. Leveraging a large public dataset of over 800 metastatic and primary samples, we identified 33 TFCGs in common between the high impact group and metastatic lesions, including SOX4/FOXA2/GATA4, and a TFCG containing JUN, JUNB, JUND, FOS, FOSB, and FOSL1. The majority of metastatic TFCGs were subsets of larger TFCGs in the high impact group network, suggesting a refinement of critical TFCGs in prostate cancer progression. CONCLUSIONS: We have identified TFCGs associated with pronounced initial transcriptional response to ADT, aggressive signatures, and metastasis. Our findings suggest multiple new hypotheses that could lead to novel combination therapies to prevent the development of CRPC following ADT.

MEDICI: Mining Essentiality Data to Identify Critical Interactions for Cancer Drug Target Discovery and Development.

Protein-protein interactions (PPIs) mediate the transmission and regulation of oncogenic signals that are essential to cellular proliferation and survival, and thus represent potential targets for anti-cancer therapeutic discovery. Despite their significance, there is no method to experimentally disrupt and interrogate the essentiality of individual endogenous PPIs. The ability to computationally predict or infer PPI essentiality would help prioritize PPIs for drug discovery and help advance understanding of cancer biology. Here we introduce a computational method (MEDICI) to predict PPI essentiality by combining gene knockdown studies with network models of protein interaction pathways in an analytic framework. Our method uses network topology to model how gene silencing can disrupt PPIs, relating the unknown essentialities of individual PPIs to experimentally observed protein essentialities. This model is then deconvolved to recover the unknown essentialities of individual PPIs. We demonstrate the validity of our approach via prediction of sensitivities to compounds based on PPI essentiality and differences in essentiality based on genetic mutations. We further show that lung cancer patients have improved overall survival when specific PPIs are no longer present, suggesting that these PPIs may be potentially new targets for therapeutic development. Software is freely available at https://github.com/cooperlab/MEDICI. Datasets are available at https://ctd2.nci.nih.gov/dataPortal.

Next generation sequencing technology and genomewide data analysis: Perspectives for retinal research.

The advent of high throughput next generation sequencing (NGS) has accelerated the pace of discovery of disease-associated genetic variants and genomewide profiling of expressed sequences and epigenetic marks, thereby permitting systems-based analyses of ocular development and disease. Rapid evolution of NGS and associated methodologies presents significant challenges in acquisition, management, and analysis of large data sets and for extracting biologically or clinically relevant information. Here we illustrate the basic design of commonly used NGS-based methods, specifically whole exome sequencing, transcriptome, and epigenome profiling, and provide recommendations for data analyses. We briefly discuss systems biology approaches for integrating multiple data sets to elucidate gene regulatory or disease networks. While we provide examples from the retina, the NGS guidelines reviewed here are applicable to other tissues/cell types as well.

Kruppel-like factor 7 overexpression suppresses hematopoietic stem and progenitor cell function.

Increased expression of Kruppel-like factor 7 (KLF7) is an independent predictor of poor outcome in pediatric acute lymphoblastic leukemia. The contribution of KLF7 to hematopoiesis has not been previously described. Herein, we characterized the effect on murine hematopoiesis of the loss of KLF7 and enforced expression of KLF7. Long-term multilineage engraftment of Klf7(-/-) cells was comparable with control cells, and self-renewal, as assessed by serial transplantation, was not affected. Enforced expression of KLF7 results in a marked suppression of myeloid progenitor cell growth and a loss of short- and long-term repopulating activity. Interestingly, enforced expression of KLF7, although resulting in multilineage growth suppression that extended to hematopoietic stem cells and common lymphoid progenitors, spared T cells and enhanced the survival of early thymocytes. RNA expression profiling of KLF7-overexpressing hematopoietic progenitors identified several potential target genes mediating these effects. Notably, the known KLF7 target Cdkn1a (p21(Cip1/Waf1)) was not induced by KLF7, and loss of CDKN1A does not rescue the repopulating defect. These results suggest that KLF7 is not required for normal hematopoietic stem and progenitor function, but increased expression, as seen in a subset of lymphoid leukemia, inhibits myeloid cell proliferation and promotes early thymocyte survival.