Exploring machine learning strategies for predicting cardiovascular disease risk factors from multi-omic data.

BACKGROUND: Machine learning (ML) classifiers are increasingly used for predicting cardiovascular disease (CVD) and related risk factors using omics data, although these outcomes often exhibit categorical nature and class imbalances. However, little is known about which ML classifier, omics data, or upstream dimension reduction strategy has the strongest influence on prediction quality in such settings. Our study aimed to illustrate and compare different machine learning strategies to predict CVD risk factors under different scenarios. METHODS: We compared the use of six ML classifiers in predicting CVD risk factors using blood-derived metabolomics, epigenetics and transcriptomics data. Upstream omic dimension reduction was performed using either unsupervised or semi-supervised autoencoders, whose downstream ML classifier performance we compared. CVD risk factors included systolic and diastolic blood pressure measurements and ultrasound-based biomarkers of left ventricular diastolic dysfunction (LVDD; E/e' ratio, E/A ratio, LAVI) collected from 1,249 Finnish participants, of which 80% were used for model fitting. We predicted individuals with low, high or average levels of CVD risk factors, the latter class being the most common. We constructed multi-omic predictions using a meta-learner that weighted single-omic predictions. Model performance comparisons were based on the F1 score. Finally, we investigated whether learned omic representations from pre-trained semi-supervised autoencoders could improve outcome prediction in an external cohort using transfer learning. RESULTS: Depending on the ML classifier or omic used, the quality of single-omic predictions varied. Multi-omics predictions outperformed single-omics predictions in most cases, particularly in the prediction of individuals with high or low CVD risk factor levels. Semi-supervised autoencoders improved downstream predictions compared to the use of unsupervised autoencoders. In addition, median gains in Area Under the Curve by transfer learning compared to modelling from scratch ranged from 0.09 to 0.14 and 0.07 to 0.11 units for transcriptomic and metabolomic data, respectively. CONCLUSIONS: By illustrating the use of different machine learning strategies in different scenarios, our study provides a platform for researchers to evaluate how the choice of omics, ML classifiers, and dimension reduction can influence the quality of CVD risk factor predictions.

MRI-Targeted Prostate Biopsy Introduces Grade Inflation and Overtreatment.

Purpose: The use of MRI-targeted biopsies has led to lower detection of Gleason Grade Group 1 (GG1) prostate cancer and increased detection of GG2 disease. Although this finding is generally attributed to improved sensitivity and specificity of MRI for aggressive cancers, it might also be explained by grade inflation. Our objective was to determine the likelihood of definitive treatment and risk of post-treatment recurrence for patients with GG2 cancer diagnosed using targeted biopsies relative to men with GG1 cancer diagnosed using systematic biopsies. Methods: We performed a retrospective study on a large tertiary centre registry (HUS Acamedic Datalake) to retrieve data on prostate cancer diagnosis, treatment, and cancer recurrence. We included patients with either GG1 with systematic biopsies (3317 men) or GG2 with targeted biopsies (554 men) from 1993 to 2019. We assessed the risk of curative treatment and recurrence after treatment. Kaplan-Meier survival curves were computed to assess treatment- and recurrence-free survival. Cox proportional hazards regression analysis was performed to assess the risk of posttreatment recurrence. Results: Patients with systematic biopsy detected GG1 cancer had a significantly longer median time-to-treatment (31 months) than those with targeted biopsy detected GG2 cancer (4 months, p<0.0001). The risk of recurrence after curative treatment was similar between groups with the upper bound of 95% CI, excluding an important difference (HR: 0.94, 95% CI [0.71-1.25], p=0.7). Conclusion: GG2 cancers detected by MRI-targeted biopsy are treated more aggressively than GG1 cancers detected by systematic biopsy, despite having similar oncologic risk. To prevent further overtreatment related to the MRI pathway, treatment guidelines from the pre-MRI era need to be updated to consider changes in the diagnostic pathway.

Spectral decoupling for training transferable neural networks in medical imaging.

Many neural networks for medical imaging generalize poorly to data unseen during training. Such behavior can be caused by overfitting easy-to-learn features while disregarding other potentially informative features. A recent implicit bias mitigation technique called spectral decoupling provably encourages neural networks to learn more features by regularizing the networks' unnormalized prediction scores with an L2 penalty. We show that spectral decoupling increases the networks' robustness for data distribution shifts and prevents overfitting on easy-to-learn features in medical images. To validate our findings, we train networks with and without spectral decoupling to detect prostate cancer on tissue slides and COVID-19 in chest radiographs. Networks trained with spectral decoupling achieve up to 9.5 percent point higher performance on external datasets. Spectral decoupling alleviates generalization issues associated with neural networks and can be used to complement or replace computationally expensive explicit bias mitigation methods, such as stain normalization in histological images.

Associations of PTEN and ERG with Magnetic Resonance Imaging Visibility and Assessment of Non-organ-confined Pathology and Biochemical Recurrence After Radical Prostatectomy.

BACKGROUND: Diagnosing clinically significant prostate cancer (PCa) is challenging, but may be facilitated by biomarkers and multiparametric magnetic resonance imaging (MRI). OBJECTIVE: To determine the association between biomarkers phosphatase and tensin homolog (PTEN) and ETS-related gene (ERG) with visible and invisible PCa lesions in MRI, and to predict biochemical recurrence (BCR) and non-organ-confined (non-OC) PCa by integrating clinical, MRI, and biomarker-related data. DESIGN, SETTING, AND PARTICIPANTS: A retrospective analysis of a population-based cohort of men with PCa, who underwent preoperative MRI followed by radical prostatectomy (RP) during 2014-2015 in Helsinki University Hospital (n = 346), was conducted. A tissue microarray corresponding to the MRI-visible and MRI-invisible lesions in RP specimens was constructed and stained for PTEN and ERG. OUTCOME MEASUREMENTS AND STATISTICAL ANALYSIS: Associations of PTEN and ERG with MRI-visible and MRI-invisible lesions were examined (Pearson's χ2 test), and predictions of non-OC disease together with clinical and MRI parameters were determined (area under the receiver operating characteristic curve and logistic regression analyses). BCR prediction was analyzed by Kaplan-Meier and Cox proportional hazard analyses. RESULTS AND LIMITATIONS: Patients with MRI-invisible lesions (n = 35) had less PTEN loss and ERG-positive expression compared with patients (n = 90) with MRI-visible lesions (17.2% vs 43.3% [p = 0.006]; 8.6% vs 20.0% [p = 0.125]). Patients with invisible lesions had better, but not statistically significantly improved, BCR-free survival probability in Kaplan-Meier analyses (p = 0.055). Rates of BCR (5.7% vs 21.1%; p = 0.039), extraprostatic extension (11.4% vs 44.6%; p < 0.001), seminal vesicle invasion (0% vs 21.1%; p = 0.003), and lymph node metastasis (0% vs 12.2%; p = 0.033) differed between the groups in favor of patients with MRI-invisible lesions. Biomarkers had no independent role in predicting non-OC disease or BCR. The short follow-up period was a limitation. CONCLUSIONS: PTEN loss, BCR, and non-OC RP findings were more often encountered with MRI-visible lesions. PATIENT SUMMARY: Magnetic resonance imaging (MRI) of the prostate misses some cancer lesions. MRI-invisible lesions seem to be less aggressive than MRI-visible lesions.

Prostate MRI added to CAPRA, MSKCC and Partin cancer nomograms significantly enhances the prediction of adverse findings and biochemical recurrence after radical prostatectomy.

BACKGROUND: To determine the added value of preoperative prostate multiparametric MRI (mpMRI) supplementary to clinical variables and their role in predicting post prostatectomy adverse findings and biochemically recurrent cancer (BCR). METHODS: All consecutive patients treated at HUS Helsinki University Hospital with robot assisted radical prostatectomy (RALP) between 2014 and 2015 were included in the analysis. The mpMRI data, clinical variables, histopathological characteristics, and follow-up information were collected. Study end-points were adverse RALP findings: extraprostatic extension, seminal vesicle invasion, lymph node involvement, and BCR. The Memorial Sloan Kettering Cancer Center (MSKCC) nomogram, Cancer of the Prostate Risk Assessment (CAPRA) score and the Partin score were combined with any adverse findings at mpMRI. Predictive accuracy for adverse RALP findings by the regression models was estimated before and after the addition of MRI results. Logistic regression, area under curve (AUC), decision curve analyses, Kaplan-Meier survival curves and Cox proportional hazard models were used. RESULTS: Preoperative mpMRI data from 387 patients were available for analysis. Clinical variables alone, MSKCC nomogram or Partin tables were outperformed by models with mpMRI for the prediction of any adverse finding at RP. AUC for clinical parameters versus clinical parameters and mpMRI variables were 0.77 versus 0.82 for any adverse finding. For MSKCC nomogram versus MSKCC nomogram and mpMRI variables the AUCs were 0.71 and 0.78 for any adverse finding. For Partin tables versus Partin tables and mpMRI variables the AUCs were 0.62 and 0.73 for any adverse finding. In survival analysis, mpMRI-projected adverse RP findings stratify CAPRA and MSKCC high-risk patients into groups with distinct probability for BCR. CONCLUSIONS: Preoperative mpMRI improves the predictive value of commonly used clinical variables for pathological stage at RP and time to BCR. mpMRI is available for risk stratification prebiopsy, and should be considered as additional source of information to the standard predictive nomograms.