Dual adversarial deconfounding autoencoder for joint batch-effects removal from multi-center and multi-scanner radiomics data.

Medical imaging represents the primary tool for investigating and monitoring several diseases, including cancer. The advances in quantitative image analysis have developed towards the extraction of biomarkers able to support clinical decisions. To produce robust results, multi-center studies are often set up. However, the imaging information must be denoised from confounding factors-known as batch-effect-like scanner-specific and center-specific influences. Moreover, in non-solid cancers, like lymphomas, effective biomarkers require an imaging-based representation of the disease that accounts for its multi-site spreading over the patient's body. In this work, we address the dual-factor deconfusion problem and we propose a deconfusion algorithm to harmonize the imaging information of patients affected by Hodgkin Lymphoma in a multi-center setting. We show that the proposed model successfully denoises data from domain-specific variability (p-value < 0.001) while it coherently preserves the spatial relationship between imaging descriptions of peer lesions (p-value = 0), which is a strong prognostic biomarker for tumor heterogeneity assessment. This harmonization step allows to significantly improve the performance in prognostic models with respect to state-of-the-art methods, enabling building exhaustive patient representations and delivering more accurate analyses (p-values < 0.001 in training, p-values < 0.05 in testing). This work lays the groundwork for performing large-scale and reproducible analyses on multi-center data that are urgently needed to convey the translation of imaging-based biomarkers into the clinical practice as effective prognostic tools. The code is available on GitHub at this https://github.com/LaraCavinato/Dual-ADAE .

A Deep Survival EWAS approach estimating risk profile based on pre-diagnostic DNA methylation: An application to breast cancer time to diagnosis.

Previous studies for cancer biomarker discovery based on pre-diagnostic blood DNA methylation (DNAm) profiles, either ignore the explicit modeling of the Time To Diagnosis (TTD), or provide inconsistent results. This lack of consistency is likely due to the limitations of standard EWAS approaches, that model the effect of DNAm at CpG sites on TTD independently. In this work, we aim to identify blood DNAm profiles associated with TTD, with the aim to improve the reliability of the results, as well as their biological meaningfulness. We argue that a global approach to estimate CpG sites effect profile should capture the complex (potentially non-linear) relationships interplaying between sites. To prove our concept, we develop a new Deep Learning-based approach assessing the relevance of individual CpG Islands (i.e., assigning a weight to each site) in determining TTD while modeling their combined effect in a survival analysis scenario. The algorithm combines a tailored sampling procedure with DNAm sites agglomeration, deep non-linear survival modeling and SHapley Additive exPlanations (SHAP) values estimation to aid robustness of the derived effects profile. The proposed approach deals with the common complexities arising from epidemiological studies, such as small sample size, noise, and low signal-to-noise ratio of blood-derived DNAm. We apply our approach to a prospective case-control study on breast cancer nested in the EPIC Italy cohort and we perform weighted gene-set enrichment analyses to demonstrate the biological meaningfulness of the obtained results. We compared the results of Deep Survival EWAS with those of a traditional EWAS approach, demonstrating that our method performs better than the standard approach in identifying biologically relevant pathways.

Development of a method for generating SNP interaction-aware polygenic risk scores for radiotherapy toxicity.

AIM: To identify the effect of single nucleotide polymorphism (SNP) interactions on the risk of toxicity following radiotherapy (RT) for prostate cancer (PCa) and propose a new method for polygenic risk score incorporating SNP-SNP interactions (PRSi). MATERIALS AND METHODS: Analysis included the REQUITE PCa cohort that received external beam RT and was followed for 2 years. Late toxicity endpoints were: rectal bleeding, urinary frequency, haematuria, nocturia, decreased urinary stream. Among 43 literature-identified SNPs, the 30% most strongly associated with each toxicity were tested. SNP-SNP combinations (named SNP-allele sets) seen in ≥10% of the cohort were condensed into risk (RS) and protection (PS) scores, respectively indicating increased or decreased toxicity risk. Performance of RS and PS was evaluated by logistic regression. RS and PS were then combined into a single PRSi evaluated by area under the receiver operating characteristic curve (AUC). RESULTS: Among 1,387 analysed patients, toxicity rates were 11.7% (rectal bleeding), 4.0% (urinary frequency), 5.5% (haematuria), 7.8% (nocturia) and 17.1% (decreased urinary stream). RS and PS combined 8 to 15 different SNP-allele sets, depending on the toxicity endpoint. Distributions of PRSi differed significantly in patients with/without toxicity with AUCs ranging from 0.61 to 0.78. PRSi was better than the classical summed PRS, particularly for the urinary frequency, haematuria and decreased urinary stream endpoints. CONCLUSIONS: Our method incorporates SNP-SNP interactions when calculating PRS for radiotherapy toxicity. Our approach is better than classical summation in discriminating patients with toxicity and should enable incorporating genetic information to improve normal tissue complication probability models.

A Deep Learning Approach Validates Genetic Risk Factors for Late Toxicity After Prostate Cancer Radiotherapy in a REQUITE Multi-National Cohort.

Background: REQUITE (validating pREdictive models and biomarkers of radiotherapy toxicity to reduce side effects and improve QUalITy of lifE in cancer survivors) is an international prospective cohort study. The purpose of this project was to analyse a cohort of patients recruited into REQUITE using a deep learning algorithm to identify patient-specific features associated with the development of toxicity, and test the approach by attempting to validate previously published genetic risk factors. Methods: The study involved REQUITE prostate cancer patients treated with external beam radiotherapy who had complete 2-year follow-up. We used five separate late toxicity endpoints: ≥grade 1 late rectal bleeding, ≥grade 2 urinary frequency, ≥grade 1 haematuria, ≥ grade 2 nocturia, ≥ grade 1 decreased urinary stream. Forty-three single nucleotide polymorphisms (SNPs) already reported in the literature to be associated with the toxicity endpoints were included in the analysis. No SNP had been studied before in the REQUITE cohort. Deep Sparse AutoEncoders (DSAE) were trained to recognize features (SNPs) identifying patients with no toxicity and tested on a different independent mixed population including patients without and with toxicity. Results: One thousand, four hundred and one patients were included, and toxicity rates were: rectal bleeding 11.7%, urinary frequency 4%, haematuria 5.5%, nocturia 7.8%, decreased urinary stream 17.1%. Twenty-four of the 43 SNPs that were associated with the toxicity endpoints were validated as identifying patients with toxicity. Twenty of the 24 SNPs were associated with the same toxicity endpoint as reported in the literature: 9 SNPs for urinary symptoms and 11 SNPs for overall toxicity. The other 4 SNPs were associated with a different endpoint. Conclusion: Deep learning algorithms can validate SNPs associated with toxicity after radiotherapy for prostate cancer. The method should be studied further to identify polygenic SNP risk signatures for radiotherapy toxicity. The signatures could then be included in integrated normal tissue complication probability models and tested for their ability to personalize radiotherapy treatment planning.

Data mining application to healthcare fraud detection: a two-step unsupervised clustering method for outlier detection with administrative databases.

BACKGROUND: The healthcare sector is an interesting target for fraudsters. The availability of a great amount of data makes it possible to tackle this issue with the adoption of data mining techniques, making the auditing process more efficient and effective. This research has the objective of developing a novel data mining model devoted to fraud detection among hospitals using Hospital Discharge Charts (HDC) in Administrative Databases. In particular, it is focused on the DRG upcoding practice, i.e., the tendency of registering codes for provided services and inpatients health status so to make the hospitalization fall within a more remunerative DRG class. METHODS: We propose a two-step algorithm: the first step entails kmeans clustering of providers to identify locally consistent and locally similar groups of hospitals, according to their characteristics and behavior treating a specific disease, in order to spot outliers within this groups of peers. An initial grid search for the best number of features to be selected (through Principal Feature Analysis) and the best number of local groups makes the algorithm extremely flexible. In the second step, we propose a human-decision support system that helps auditors cross-validating the identified outliers, analyzing them w.r.t. fraud-related variables, and the complexity of patients' casemix they treated. The proposed algorithm was tested on a database relative to HDC collected by Regione Lombardia (Italy) in a time period of three years (2013-2015), focusing on the treatment of Heart Failure. RESULTS: The model identified 6 clusters of hospitals and 10 outliers among the 183 units. Out of those providers, we report the in depth the application of Step Two on three Hospitals (two private and one public). Cross-validating with the patients' population and the hospitals' characteristics, the public hospital seemed justified in its outlierness, while the two private providers were deemed interesting for a further investigation by auditors. CONCLUSIONS: The proposed model is promising in identifying anomalous DRG coding behavior and it is easily transferrable to all diseases and contexts of interest. Our proposal contributes to the limited literature regarding behavioral models for fraud detection, identifying the most 'cautious' fraudsters. The results of the first and the second Steps together represent a valuable set of information for auditors in their preliminary investigation.