Faculty Bibliography

Objective: To determine whether a machine learning model can predict patients at risk for shoulder dystocia (SD) better than estimations of birthweight (BW) alone.
Study Design: This was a retrospective analysis of 17,731 pregnant individuals from 7/2013 to 10/2018. Utilizing a machine learning model, a total of 122 binary and continuous variables were included. Baseline models were built with different sets of variables during 3 time periods: 57 in antepartum period, 96 in stage 1 and 119 in stage 2 of labor. BW was used as a proxy for estimated fetal weight (EFW) because documented assessment of EFW was not available in all cases. Two decision tree-based models, Random Forest and XGBoost, were used as predictive models and performance was evaluated with 5-fold cross validation. Area under the receiver operating characteristic curve (AUROC) and area under precision/recall curve (AUPR) were used as metrics for evaluating model performance. Mean and standard error of performance metrics were calculated.
Result(s): The cohort included 21,232 vaginal deliveries. There were 415 (1.95%) vaginal deliveries complicated by SD that occurred in 406 patients. An AUROC of 0.73 +/- 0.01 (mean +/- standard error) was achieved for RandomForest model and 0.78 +/- 0.01 for XGBoost model for BW. BW was then added as a variable to variable sets from the antepartum period and each labor stage to assess performance change. The RandomForest model predicted patients at risk for SD better than BW alone, but only the AUROC was statistically significant at all stages (p < 0.05). (Figure 1) This finding was not demonstrated in the XGBoost model.
Conclusion(s): Our machine learning model performed better in predicting SD than EFW alone (using BW as a proxy) at each time period evaluated - the antepartum period, 1st stage and 2nd stage of labor. While our results indicate that our model may enhance the prediction of SD, these findings should be validated using a more robust data set that includes documented EFW to account for the margin of error between BW and EFW. [Formula presented] [Formula presented]
Copyright

PURPOSE/OBJECTIVE:Timely and appropriate discharge placement for patients who have undergone radical cystectomy (RC) remains challenging. Our objective was to improve the discharge planning process by creating a machine learning model that helps to predict the need for non-home hospital discharge to a higher level of care. MATERIALS AND METHODS/METHODS:Patients undergoing elective radical cystectomy for bladder cancer from 2014-2019 were identified in the ACS-NSQIP database. A gradient boosted decision tree was trained on selected predischarge variables to predict discharge location, dichotomized into home and non-home. We used threshold-moving to calibrate model predictions and evaluated model performance on a testing set using receiver operating characteristic and precision recall curves. Model performance was further examined in subgroups of interest. RESULTS AND CONCLUSIONS/CONCLUSIONS:A total of 11,881 patients met inclusion criteria with a mean age of 68.6 years. 10.6% of patients undergoing RC had non-home discharges. Our model predicting non-home discharge achieved an area under the receiver operating characteristic curve of 0.80 and an average precision of 0.33. After threshold-moving, our model had a recall of 0.757 and a precision of 0.211. Top variables by importance were septic shock occurrence, ventilator-use greater than 48 hours, organ space surgical site infection and unplanned intubation. Our model shows strong performance in identifying patients who required non-home discharge to higher levels of care, outperforming commonly used clinical indices and prior work. Modern machine learning techniques may be applied to support more timely and appropriate clinical decision making.

How cells control gene expression is a fundamental question. The relative contribution of protein-level and RNA-level regulation to this process remains unclear. Here, we perform a proteogenomic analysis of tumors and untransformed cells containing somatic copy number alterations (SCNAs). By revealing how cells regulate RNA and protein abundances of genes with SCNAs, we provide insights into the rules of gene regulation. Protein complex genes have a strong protein-level regulation while non-complex genes have a strong RNA-level regulation. Notable exceptions are plasma membrane protein complex genes, which show a weak protein-level regulation and a stronger RNA-level regulation. Strikingly, we find a strong negative association between the degree of RNA-level and protein-level regulation across genes and cellular pathways. Moreover, genes participating in the same pathway show a similar degree of RNA- and protein-level regulation. Pathways including translation, splicing, RNA processing, and mitochondrial function show a stronger protein-level regulation while cell adhesion and migration pathways show a stronger RNA-level regulation. These results suggest that the evolution of gene regulation is shaped by functional constraints and that many cellular pathways tend to evolve one predominant mechanism of gene regulation at the protein level or at the RNA level.

PURPOSE:Adjuvant immunotherapy produces durable benefit for patients with resected melanoma, but many develop recurrence and/or immune-related adverse events (irAE). We investigated whether baseline serum autoantibody (autoAb) signatures predicted recurrence and severe toxicity in patients treated with adjuvant nivolumab, ipilimumab, or ipilimumab plus nivolumab. EXPERIMENTAL DESIGN:This study included 950 patients: 565 from CheckMate 238 (408 ipilimumab versus 157 nivolumab) and 385 from CheckMate 915 (190 nivolumab versus 195 ipilimumab plus nivolumab). Serum autoAbs were profiled using the HuProt Human Proteome Microarray v4.0 (CDI Laboratories, Mayaguez, PR). Analysis of baseline differentially expressed autoAbs was followed by recurrence and severe toxicity signature building for each regimen, testing of the signatures, and additional independent validation for nivolumab using patients from CheckMate 915. RESULTS:In the nivolumab independent validation cohort, high recurrence score predicted significantly worse recurrence-free survival [RFS; adjusted HR (aHR), 3.60; 95% confidence interval (CI), 1.98-6.55], and outperformed a model composed of clinical variables including PD-L1 expression (P < 0.001). Severe toxicity score was a significant predictor of severe irAEs (aHR, 13.53; 95% CI, 2.59-86.65). In the ipilimumab test cohort, high recurrence score was associated with significantly worse RFS (aHR, 3.21; 95% CI, 1.38-7.45) and severe toxicity score significantly predicted severe irAEs (aHR, 11.04; 95% CI, 3.84-37.25). In the ipilimumab plus nivolumab test cohort, high autoAb recurrence score was associated with significantly worse RFS (aHR, 6.45; 95% CI, 1.48-28.02), and high severe toxicity score was significantly associated with severe irAEs (aHR, 23.44; 95% CI, 4.10-212.50). CONCLUSIONS:Baseline serum autoAb signatures predicted recurrence and severe toxicity in patients treated with adjuvant immunotherapy. Prospective testing of the signatures that include datasets with longer follow-up and rare but more severe toxicities will help determine their generalizability and potential clinical utility. See related commentary by Hassel and Luke, p. 3914.

BACKGROUND:Postpartum hemorrhage remains one of the largest causes of maternal morbidity and mortality in the United States. OBJECTIVE:The aim of this paper is to use machine learning techniques to identify patients at risk for postpartum hemorrhage at obstetric delivery. METHODS:Women aged 18 to 55 years delivering at a major academic center from July 2013 to October 2018 were included for analysis (N=30,867). A total of 497 variables were collected from the electronic medical record including the following: demographic information; obstetric, medical, surgical, and family history; vital signs; laboratory results; labor medication exposures; and delivery outcomes. Postpartum hemorrhage was defined as a blood loss of ≥1000 mL at the time of delivery, regardless of delivery method, with 2179 (7.1%) positive cases observed. Supervised learning with regression-, tree-, and kernel-based machine learning methods was used to create classification models based upon training (21,606/30,867, 70%) and validation (4630/30,867, 15%) cohorts. Models were tuned using feature selection algorithms and domain knowledge. An independent test cohort (4631/30,867, 15%) determined final performance by assessing for accuracy, area under the receiver operating curve (AUROC), and sensitivity for proper classification of postpartum hemorrhage. Separate models were created using all collected data versus models limited to data available prior to the second stage of labor or at the time of decision to proceed with cesarean delivery. Additional models examined patients by mode of delivery. RESULTS:Gradient boosted decision trees achieved the best discrimination in the overall model. The model including all data mildly outperformed the second stage model (AUROC 0.979, 95% CI 0.971-0.986 vs AUROC 0.955, 95% CI 0.939-0.970). Optimal model accuracy was 98.1% with a sensitivity of 0.763 for positive prediction of postpartum hemorrhage. The second stage model achieved an accuracy of 98.0% with a sensitivity of 0.737. Other selected algorithms returned models that performed with decreased discrimination. Models stratified by mode of delivery achieved good to excellent discrimination but lacked the sensitivity necessary for clinical applicability. CONCLUSIONS:Machine learning methods can be used to identify women at risk for postpartum hemorrhage who may benefit from individualized preventative measures. Models limited to data available prior to delivery perform nearly as well as those with more complete data sets, supporting their potential utility in the clinical setting. Further work is necessary to create successful models based upon mode of delivery and to validate the findings of this study. An unbiased approach to hemorrhage risk prediction may be superior to human risk assessment and represents an area for future research.

A recent study by Saldanha et al. demonstrates that blockchain-based models outcompeted local models and performed similarly with merged models to predict molecular features from cancer histopathology images. The results reveal the capability of decentralized models in molecular diagnosis of cancer.

Image-based analysis as a method for mutation detection can be advantageous in settings when tumor tissue is limited or unavailable for direct testing. Here, we utilize two distinct and complementary machine learning methods of analyzing whole slide images (WSI) for predicting mutated BRAF. In the first method, WSI of melanomas from 256 patients were used to train a deep convolutional neural network (CNN) in order to develop a fully automated model that first selects for tumor-rich areas (Area Under the Curve AUC=0.96) then predicts for mutated BRAF (AUC=0.71). Saliency mapping was performed and revealed that pixels corresponding to nuclei were the most relevant to network learning. In the second method, WSI were analyzed using a pathomics pipeline that first annotates nuclei and then quantifies nuclear features, demonstrating that mutated BRAF nuclei were significantly larger and rounder nuclei compared to BRAF WT nuclei. Lastly, we developed a model that combines clinical information, deep learning, and pathomics that improves the predictive performance for mutated BRAF to AUC=0.89. Not only does this provide additional insights on how BRAF mutations affect tumor structural characteristics, machine learning-based analysis of WSI has the potential to be integrated into higher order models for understanding tumor biology.

The deubiquitinase USP1 is a critical regulator of genome integrity through the deubiquitylation of Fanconi Anemia proteins and the DNA replication processivity factor, proliferating cell nuclear antigen (PCNA). Uniquely, following UV irradiation, USP1 self-inactivates through autocleavage, which enables its own degradation and in turn, upregulates PCNA monoubiquitylation. However, the functional role for this autocleavage event during physiological conditions remains elusive. Herein, we discover that cells harboring an autocleavage-defective USP1 mutant, while still able to robustly deubiquitylate PCNA, experience more replication fork-stalling and premature fork termination events. Using super-resolution microscopy and live-cell single-molecule tracking, we show that these defects are related to the inability of this USP1 mutant to be properly recycled from sites of active DNA synthesis, resulting in replication-associated lesions. Furthermore, we find that the removal of USP1 molecules from DNA is facilitated by the DNA-dependent metalloprotease Spartan to counteract the cytotoxicity caused by "USP1-trapping". We propose a utility of USP1 inhibitors in cancer therapy based on their ability to induce USP1-trapping lesions and consequent replication stress and genomic instability in cancer cells, similar to how non-covalent DNA-protein crosslinks cause cytotoxicity by imposing steric hindrances upon proteins involved in DNA transactions.

Retrotransposons are genomic DNA sequences that copy themselves to new genomic locations via RNA intermediates; LINE-1 is the only active and autonomous retrotransposon in the human genome. The mobility of LINE-1 is largely repressed in somatic tissues but is derepressed in many cancers, where LINE-1 retrotransposition is correlated with p53 mutation and copy number alteration (CNA). In cell lines, inducing LINE-1 expression can cause double-strand breaks (DSBs) and replication stress. Reanalyzing multiomic data from breast, ovarian, endometrial, and colon cancers, we confirmed correlations between LINE-1 expression, p53 mutation status, and CNA. We observed a consistent correlation between LINE-1 expression and the abundance of DNA replication complex components, indicating that LINE-1 may also induce replication stress in human tumors. In endometrial cancer, high-quality phosphoproteomic data allowed us to identify the DSB-induced ATM-MRN-SMC S phase checkpoint pathway as the primary DNA damage response (DDR) pathway associated with LINE-1 expression. Induction of LINE-1 expression in an in vitro model led to increased phosphorylation of MRN complex member RAD50, suggesting that LINE-1 directly activates this pathway.