Machine Learning Random Forest Model Research Articles

Long-term prediction of algal chlorophyll based on empirical models and the machine learning approach in relation to trophic variation in Juam Reservoir, Korea

This study analyzed spatiotemporal variation and long-term trends in water quality indicators and trophic state conditions in an Asian temperate reservoir, Juam Reservoir (JR), and developed models that forecast algal chlorophyll (CHL-a) over a period of 30 years, 1993–2022. The analysis revealed that there were longitudinal gradients in water quality indicators along the reservoir, with notable influences from tributaries and seasonal variations in nutrient regimes and suspended solids. The empirical model showed phosphorus was found to be the key determinant of algal biomass, while suspended solids played a significant role in regulating water transparency. The trophic state indices indicated varying levels of trophic status, ranging from mesotrophic to eutrophic. Eutrophic states were particularly observed in zones after the summer monsoons, indicating a heightened risk of algal blooms, which were more prevalent in flood years. The analysis of trophic state index deviation suggested that phosphorus availability strongly influences the reservoir trophic status, with several episodes of non-algal turbidity at each site during Mon. Increases in non-algal turbidity were more prevalent during the monsoon in flood years. This study also highlighted overall long-term trends in certain water quality parameters, albeit with indications of shifting pollution sources towards non-biodegradable organic matter. According to the machine learning tests, a random forest (RF) model strongly predicted CHL-a (R2 = 0.72, p < 0.01), except for algal biomass peaks (>60 μg/L), compared to all other models. Overall, our research suggests that CHL-a and trophic variation are primarily regulated by the monsoon intensity and predicted well by the machine learning RF model.

Discordance between aPTT and anti-Xa in monitoring heparin anticoagulation in mechanical circulatory support.

It is unclear whether activated partial thromboplastin time (aPTT) or anti-Xa is more accurate for monitoring heparin anticoagulation in mechanical circulatory support (MCS) patients. This study investigates the relationship between aPTT and anti-Xa in MCS patients and identifies predictors of discordance. aPTT and anti-Xa were simultaneously measured in a prospective cohort of MCS patients receiving unfractionated heparin at a tertiary academic medical centre. Therapeutic aPTT and anti-Xa levels were 60-100s and 0.3-0.7IU/mL, respectively, and concordance was defined as both levels being subtherapeutic, therapeutic, or supratherapeutic. To identify predictors of discordance, both a machine learning random forest model and a multivariate regression model were applied to patient demographics, device type, and 14 laboratory variables; 23001 pairs of simultaneously measured aPTT/anti-Xa were collected from 699 MCS patients. aPTT and anti-Xa were concordant in 35.5% of paired observations and discordant in 64.5% (aPTT>antiXa 61.5%; aPTT<antiXa 3.0%). Discordance with a high aPTT relative to anti-Xa (aPTT>antiXa) was associated with high INR, eGFR, and total bilirubin, as well as low platelets, haemoglobin, pre-albumin, white blood cell count, and haptoglobin. Total artificial heart and durable ventricular assist devices were more likely to be associated with aPTT>anti-Xa than temporary MCS devices. aPTT and anti-Xa were frequently discordant in MCS patients receiving heparin anticoagulation. Clinical conditions common in MCS patients such as concurrent warfarin use, malnutrition, haemolysis, and thrombocytopenia, as well as durable type of MCS devices were associated with a high aPTT relative to anti-Xa.

Abstract A066: Machine learning elucidates biology of response within and outside the mechanisms of action of therapeutic agents in the I-SPY2 breast cancer TRIAL

Abstract Background: Machine learning (ML) in translational medicine has led to prediction of clinical outcomes and identification of new biomarkers. We employ ML in prediction of pathologic complete response (pCR) in high-risk breast cancer patients in the neoadjuvant I-SPY2 TRIAL where not all novel agents have strong predictive biomarkers. Leveraging a ML approach using progressively expanded candidate genes, we explore the limitations of using only known mechanisms of action in predicting pCR, and the extent to which biology outside known drug action improves response prediction in the first 10 arms of the trial. Methods: ML random forest models were developed in I-SPY2 patients (n=982) with pre-treatment gene expression and pCR data across 10 treatment arms (PMID: 35623341), including inhibitors of HER2: neratinib (N), pertuzumab (P), TDM1/P; AKT (MK-2206); IGF1R (ganitumab); HSP90 (ganetespib); PARP/DNA repair (veliparib/carboplatin, VC); ANG1/2 (trebananib, T); immune checkpoints (PD1-inh); and Control (Ctr). Each HR/HER2 receptor/treatment arm subset (m=27) was evaluated independently. We employed a three-pronged feature-selection approach using (1) genes restricted to known mechanism of action of individual I-SPY2 agents (k=10 to 88 genes); (2) genes expanded to include targeted pathways for all 10 agents/combinations (k=282); and (3) an unbiased whole genome approach (k=17,990). Samples were partitioned with 75% used for training and cross-validation, and 25% held out as test sets. Predictive ML models were defined as those with performance ≥ 0.90 based on different performance metrics (e.g., AUC, sensitivity, specificity). Results: For each of the 27 subtype-treatment subsets, at least one high performing model was identified. In 6 subtype-treatment subsets, mechanism of action genes were sufficient to predict pCR: AKT/PI3K/HER genes in HR+HER2- N and HR-HER2+ P; DNA repair genes in HR+HER2- VC; angiogenesis-associated genes in HR+HER2+ T; and immune-associated genes in both HR+HER2- and HR-HER2- PD1-inh subsets. Expanded targeted pathway models were required to identify predictive models in 8 additional subtype-treatment pairs from the N, T-DM1/P, MK-2206, VC, T, and HER2+ Ctr arms, with significant contribution of DNA repair, immune, and HSP90 genes for multiple arms. A genome-wide approach was required for the remaining 13 subtype-treatment pairs with no previous models from the N, P, MK-2206, ganitumab, ganetespib, T, and HER2- Ctr arms. Even for subtype-treatment pairs where mechanism of action gene sets was sufficient for reasonable models, expanded gene sets resulted in improved performance. For instance, metabolism genes improved model performance for HR-HER2+ in N and Ctr, and for HR+HER2- in the PD1-inh arm; and mitochondrial and protein folding dysfunction genes improved response prediction in HR-HER2- in the ganetespib arm. Conclusion: Our study identifies mechanism of action biomarkers associated with response to each drug and elucidates possible off-target effects contributing to observed drug sensitivity and resistance. Citation Format: Rosalyn W. Sayaman, Denise M. Wolf, Christina Yau, Julia Wulfkhule, Emanuel F. Petricoin, Lamorna Brown-Swigart, Tam Binh Bui, Gillian L. Hirst, Diane Heditsian, W. Fraser Symmans, Angela DeMichele, Mark LaBarge, Laura J. Esserman, Laura van ‘t Veer. Machine learning elucidates biology of response within and outside the mechanisms of action of therapeutic agents in the I-SPY2 breast cancer TRIAL [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Breast Cancer Research; 2023 Oct 19-22; San Diego, California. Philadelphia (PA): AACR; Cancer Res 2024;84(3 Suppl_1):Abstract nr A066.

Meteorologically normalized spatial and temporal variations investigation using a machine learning-random forest model in criteria pollutants across Tehran, Iran

Investigating road traffic contribution to urban air pollution is helpful to determine traffic management approaches and environmental intervention policies that focus on better air quality. To this end, hourly average values of air pollutants collected from 2015 to 2021 from 20 monitoring sites including 10 traffic (TS) and 10 urban background (UB) sites in Tehran, Iran were retrieved. A machine learning-random forest (RF) model was applied to decouple the meteorology effects of the observed air pollutant values. The meteorologically normalized concentrations of air quality data were used for calculating the traffic increment and trend analysis. The seasonal component showed the greatest importance for predicting PM2.5 and PM10 and boundary layer heights (BLH) was the dominant explanatory variable for the predictions of CO, NO, NO2, and O3 concentrations. The overall annual average value of traffic increment during study period was found as PM10 (10.5 μg m−3), PM2.5 (0.19 μg m−3), NO (9.24 ppb), NO2 (3.67 ppb), NOx (15.6 ppb), and SO2 (0.33 ppb). Ozone indicates a traffic decrement in concentration (−0.70 ppb). The Theil-Sen estimated slopes showed meteorological normalized negative trends for PM2.5, PM10, CO, NO, NOx, and SO2, while meteorological normalized positive trends were observed for NO2 and O3. The PM10, CO, NO, NOx, and SO2 concentrations trends at the traffic sites that experience more anthropogenic emissions revealed greater reductions when emission interventions or control policies are used. The higher increase of NO2 at the UB sites than the traffic sites can be due to the weaker control policies, high traffic volume, higher number of diesel-fueled vehicles, and a higher NO to NO2 conversion because of the relatively high levels of temperature and O3. This quantitative summary provides useful information for decision-makers for evaluating the intervention policies related to road traffic pollutants.

Metagenomics-Based Microbial Ecological Community Threshold and Indicators of Anthropogenic Disturbances in Estuarine Sediments.

Assessing the impacts of cumulative anthropogenic disturbances on estuarine ecosystem health is challenging. Using spatially distributed sediments from the Pearl River Estuary (PRE) in southern China, which are significantly influenced by anthropogenic activities, we demonstrated that metagenomics-based surveillance of benthic microbial communities is a robust approach to assess anthropogenic impacts on estuarine benthic ecosystems. Correlational and threshold analyses between microbial compositions and environmental conditions indicated that anthropogenic disturbances in the PRE sediments drove the taxonomic and functional variations in the benthic microbial communities. An ecological community threshold of anthropogenic disturbances was identified, which delineated the PRE sediments into two groups (H and L) with distinct taxa and functional traits. Group H, located nearshore and subjected to a higher level of anthropogenic disturbances, was enriched with pollutant degraders, putative human pathogens, fecal pollution indicators, and functional traits related to stress tolerance. In contrast, Group L, located offshore and subjected to a lower level of anthropogenic disturbances, was enriched with halotolerant and oligotrophic taxa and functional traits related to growth and resource acquisition. The machine learning random forest model identified a number of taxonomic and functional indicators that could differentiate PRE sediments between Groups H and L. The identified ecological community threshold and microbial indicators highlight the utility of metagenomics-based microbial surveillance in assessing the adverse impacts of anthropogenic disturbances in estuarine sediments, which can assist environmental management to better protect ecosystem health.

Dynamics of the Gut Mycobiome in Patients With Ulcerative Colitis

Intestinal fungi have been implicated in the pathogenesis of ulcerative colitis (UC). However, it remains unclear if fungal composition is altered during active versus quiescent disease. We analyzed clinical and metagenomic data from the Study of a Prospective Adult Research Cohort with Inflammatory Bowel Disease (SPARC IBD), available via the IBD Plexus Program of the Crohn's & Colitis Foundation. We evaluated the fungal composition of fecal samples from 421 patients with UC during clinical activity and remission. Within a longitudinal subcohort (n= 52), we assessed for dynamic taxonomic changes across alterations in clinical activity over time. We examined if fungal amplicon sequence variants and fungal-bacterial relationships were altered during activity versus remission. Finally, we classified activity in UC using a supervised machine learning random forest model trained on fungal abundance data. During clinical activity, the relative abundance of genus Candida was increased 3.5-fold (P-adj < 1× 10-4) compared with during remission. Patients with longitudinal reductions in clinical activity demonstrated parallel reductions in Candida relative abundance (P < .05). Candida relative abundance correlated with Parabacteroides diastonis, Faecalibacterium prausnitzii, and Bacteroides dorei relative abundance (P < .05) during remission; however, these correlations were disrupted during activity. Fungal abundance data successfully classified patients with active or quiescent UC (area under the curve ∼0.80), with Candida relative abundance critical to the success of the model. Clinical activity in UC is associated with an increased relative abundance of Candida, cross-sectionally and dynamically over time. The role of fecal Candida as a target for therapeutics in UC should be evaluated.

Establishment and validation of a clinical nomogram model based on serum YKL-40 to predict major adverse cardiovascular events during hospitalization in patients with acute ST-segment elevation myocardial infarction.

This study aimed to investigate the predictive value of a clinical nomogram model based on serum YKL-40 for major adverse cardiovascular events (MACE) during hospitalization in patients with acute ST-segment elevation myocardial infarction (STEMI). In this study, 295 STEMI patients from October 2020 to March 2023 in the Second People's Hospital of Hefei were randomly divided into a training group (n = 206) and a validation group (n = 89). Machine learning random forest model was used to select important variables and multivariate logistic regression was included to analyze the influencing factors of in-hospital MACE in STEMI patients; a nomogram model was constructed and the discrimination, calibration, and clinical effectiveness of the model were verified. According to the results of random forest and multivariate analysis, we identified serum YKL-40, albumin, blood glucose, hemoglobin, LVEF, and uric acid as independent predictors of in-hospital MACE in STEMI patients. Using the above parameters to establish a nomogram, the model C-index was 0.843 (95% CI: 0.79-0.897) in the training group; the model C-index was 0.863 (95% CI: 0.789-0.936) in the validation group, with good predictive power; the AUC (0.843) in the training group was greater than the TIMI risk score (0.648), p < 0.05; and the AUC (0.863) in the validation group was greater than the TIMI risk score (0.795). The calibration curve showed good predictive values and observed values of the nomogram; the DCA results showed that the graph had a high clinical application value. In conclusion, we constructed and validated a nomogram based on serum YKL-40 to predict the risk of in-hospital MACE in STEMI patients. This model can provide a scientific reference for predicting the occurrence of in-hospital MACE and improving the prognosis of STEMI patients.

Prognostic MicroRNA Fingerprints Predict Recurrence of Early-Stage Hepatocellular Carcinoma Following Hepatectomy.

Purpose: This study aims to develop liquid biopsy assays for early HCC diagnosis and prognosis. Methods: Twenty-three microRNAs were first consolidated as a panel (HCCseek-23 panel) based on their reported functions in HCC development. Serum samples were collected from 103 early-stage HCC patients before and after hepatectomy. Quantitative PCR and machine learning random forest models were applied to develop diagnostic and prognostic models. Results: For HCC diagnosis, HCCseek-23 panel demonstrated 81% sensitivity and 83% specificity for identifying HCC in the early-stage; it showed 93% sensitivity for identifying alpha-fetoprotein (AFP)-negative HCC. For HCC prognosis, the differential expressions of 8 microRNAs (HCCseek-8 panel: miR-145, miR-148a, miR-150, miR-221, miR-223, miR-23a, miR-374a, and miR-424) were significantly associated with disease-free survival (DFS) (Log-rank test p-value = 0.001). Further model improvement using these HCCseek-8 panel in combination with serum biomarkers (i.e. AFP, ALT, and AST) demonstrated a significant association with DFS (Log-rank p-value = 0.011 and Cox proportional hazards analyses p-value = 0.002). Conclusion: To the best of our knowledge, this is the first report to integrate circulating miRNAs, AST, ALT, AFP, and machine learning for predicting DFS in early HCC patients undergoing hepatectomy. In this setting, HCCSeek-23 panel is a promising circulating microRNA assay for diagnosis, while HCCSeek-8 panel is promising for prognosis to identify early HCC recurrence.

Nitrogen isotope enrichment predicts growth response of Pinus radiata in New Zealand to nitrogen fertiliser addition

The fertiliser growth response of planted forests can vary due to differences in site-specific factors like climate and soil fertility. We identified when forest stands responded to a standard, single application of nitrogen (N) fertiliser and employed a machine learning random forest model to test the use of natural abundance stable isotopic N (δ15N) to predict site response. Pinus radiata growth response was calculated as the change in periodic annual increment of basal area (PAI BA) from replicated control and treatment (~ 200 kg N ha−1) plots within trials across New Zealand. Variables in the analysis were climate, silviculture, soil, and foliage chemical properties, including natural abundance δ15N values as integrators of historical patterns in N cycling. Our Random Forest model explained 78% of the variation in growth with tree age and the δ15N enrichment factor (δ15Nfoliage − δ15Nsoil) showing more than 50% relative importance to the model. Tree growth rates generally decreased with more negative δ15N enrichment factors. Growth response to N fertiliser was highly variable. If a response was going to occur, it was most likely within 1–3 years after fertiliser addition. The Random Forest model predicts that younger stands (< 15 years old) with the freedom to grow and sites with more negative δ15N isotopic enrichment factors will exhibit the biggest growth response to N fertiliser. Supporting the challenge of forest nutrient management, these findings provide a novel decision-support tool to guide the intensification of nutrient additions.

A simplified prediction model for end-stage kidney disease in patients with diabetes

This study aimed to develop a simplified model for predicting end-stage kidney disease (ESKD) in patients with diabetes. The cohort included 2549 individuals who were followed up at Kyushu University Hospital (Japan) between January 1, 2008 and December 31, 2018. The outcome was a composite of ESKD, defined as an eGFR < 15 mL min−1 [1.73 m]−2, dialysis, or renal transplantation. The mean follow-up was 5.6 pm 3.7 years, and ESKD occurred in 176 (6.2%) individuals. Both a machine learning random forest model and a Cox proportional hazard model selected eGFR, proteinuria, hemoglobin A1c, serum albumin levels, and serum bilirubin levels in a descending order as the most important predictors among 20 baseline variables. A model using eGFR, proteinuria and hemoglobin A1c showed a relatively good performance in discrimination (C-statistic: 0.842) and calibration (Nam and D’Agostino chi2 statistic: 22.4). Adding serum albumin and bilirubin levels to the model further improved it, and a model using 5 variables showed the best performance in the predictive ability (C-statistic: 0.895, chi2 statistic: 7.7). The accuracy of this model was validated in an external cohort (n = 5153). This novel simplified prediction model may be clinically useful for predicting ESKD in patients with diabetes.

Disturbance of serum lipid metabolites and potential biomarkers in the Bleomycin model of pulmonary fibrosis in young mice

BackgroundAltered metabolic pathways have recently been considered as potential drivers of idiopathic pulmonary fibrosis (IPF) for the study of drug therapeutic targets. However, our understanding of the metabolite profile during IPF formation is lacking.MethodsTo comprehensively characterize the metabolic disorders of IPF, a mouse IPF model was constructed by intratracheal injection of bleomycin into C57BL/6J male mice, and lung tissues from IPF mice at 7 days, 14 days, and controls were analyzed by pathology, immunohistochemistry, and Western Blots. Meanwhile, serum metabolite detections were conducted in IPF mice using LC–ESI–MS/MS, KEGG metabolic pathway analysis was applied to the differential metabolites, and biomarkers were screened using machine learning algorithms.ResultsWe analyzed the levels of 1465 metabolites and found that more than one-third of the metabolites were altered during IPF formation. There were 504 and 565 metabolites that differed between M7 and M14 and controls, respectively, while 201 differential metabolites were found between M7 and M14. In IPF mouse sera, about 80% of differential metabolite expression was downregulated. Lipids accounted for more than 80% of the differential metabolite species with down-regulated expression. The KEGG pathway enrichment analysis of differential metabolites was mainly enriched to pathways such as the metabolism of glycerolipids and glycerophospholipids. Eight metabolites were screened by a machine learning random forest model, and receiver operating characteristic curves (ROC) assessed them as ideal diagnostic tools.ConclusionsIn conclusion, we have identified disturbances in serum lipid metabolism associated with the formation of pulmonary fibrosis, contributing to the understanding of the pathogenesis of pulmonary fibrosis.

Yeast and Lactic Acid Bacteria Dominate the Core Microbiome of Fermented ‘Hairy’ Tofu (Mao Tofu)

The process of fermenting tofu extends back thousands of years and is an indispensable part of Chinese culture. Despite a cultural resurgence in fermented foods and interest in microbiomes, there is little knowledge on the microbial diversity represented in fermented ‘hairy’ tofu, known locally in China as Mao tofu. High-throughput metagenomic sequencing of the ITS, LSU and 16S rDNA was used to determine Mao tofu’s fungal and bacterial community diversity across four wet markets in Yunnan, China. The results show that hairy tofu in this region consists of around 170 fungal and 365 bacterial taxa, and that microbial taxa differ between markets. Diversity also differed based on the specific niche of the tofu block, comparing the outside rind-like niche to that of the inside of the tofu block. Machine learning random forest models were able to accurately classify both the market and niche of sample origin. An over-abundance of yeast and Geotrichum was found, and Mucor (Mucoromycota) was abundant in the outside rind-like niche, which consists of the visible ‘hairy’ mycelium. The majority of the bacterial OTUs belonged to Proteobacteria, Firmicutes, and Bacteroidetes, with Acinetobacter, Lactobacillus, Sphingobacterium and Flavobacterium the most abundant genera. Putative fungal pathogens of plants (Cercospora, Diaporthe, Fusarium) and animals (Metarhizium, Entomomortierella, Pyxidiophora, Candida, Clavispora) were also detected, as were putative bacterial pathogens identified as Legionella. Non-fungal eukaryotic taxa detected by LSU amplicon sequencing included soybean (Glycine max), Protozoa, Metazoa (e.g., Nematoda and Platyhelminthes), Rhizaria and Chromista, indicating that additional biodiversity exists in the hairy tofu microbiome.

Spatiotemporal Monitoring of Soil CO2 Efflux in a Subtropical Forest during the Dry Season Based on Field Observations and Remote Sensing Imagery

The CO2 efflux from forest soil (FCO2) is one of the largest components of the global carbon cycle. Accurate estimation of FCO2 can help us better understand the carbon cycle in forested areas and precisely predict future climate change. However, the scarcity of field-measured FCO2 data in the subtropical forested area greatly limits our understanding of FCO2 dynamics at regional and global scales. This study used an automatic cavity ring-down spectrophotometer (CRDS) analyzer to measure FCO2 in a typical subtropical forest of southern China in the dry season. We found that the measured FCO2 at two experimental areas experienced similar temporal trends in the dry season and reached the minima around December, whereas the mean FCO2 differed apparently across the two areas (9.05 vs. 5.03 g C m−2 day−1) during the dry season. Moreover, we found that both abiotic (soil temperature and moisture) and biotic (vegetation productivity) factors are significantly and positively correlated, respectively, with the FCO2 variation during the study period. Furthermore, a machine-learning random forest model (RF model) that incorporates remote sensing data is developed and used to predict the FCO2 pattern in the subtropical forest, and the topographic effects on spatiotemporal patterns of FCO2 were further investigated. The model evaluation indicated that the proposed model illustrated high prediction accuracy for the training and testing dataset. Based on the proposed model, the spatiotemporal patterns of FCO2 in the forested watershed that encloses the two monitoring sites were mapped. Results showed that the spatial distribution of FCO2 is obviously affected by topography: the high FCO2 values mainly occur in relatively high altitudinal areas, in slopes of 10–25°, and in sunny slopes. The results emphasized that future studies should consider topographical effects when simulating FCO2 in subtropical forests. Overall, our study unraveled the spatiotemporal variations of FCO2 and their driving factors in a subtropical forest of southern China in the dry season, and demonstrated that the proposed RF model in combination with remote sensing data can be a useful tool for predicting FCO2 in forested areas, particularly in subtropical and tropical forest ecosystems.

Applying a machine learning modelling framework to predict delayed linkage to care in patients newly diagnosed with HIV in Mecklenburg County, North Carolina, USA.

Machine learning has the potential to help researchers better understand and close the gap in HIV care delivery in large metropolitan regions such as Mecklenburg County, North Carolina, USA. We aim to identify important risk factors associated with delayed linkage to care for HIV patients with novel machine learning models and identify high-risk regions of the delay. Deidentified 2013-2017 Mecklenburg County surveillance data in eHARS format were requested. Both univariate analyses and machine learning random forest model (developed in R 3.5.0) were applied to quantify associations between delayed linkage to care (>30 days after diagnosis) and various risk factors for individual HIV patients. We also aggregated linkage to care by zip codes to identify high-risk communities within the county. Types of HIV-diagnosing facility significantly influenced time to linkage; first diagnosis in hospital was associated with the shortest time to linkage. HIV patients with lower CD4+ cell counts (<200/ml) were twice as likely to link to care within 30 days than those with higher CD4+ cell count. Random forest model achieved high accuracy (>80% without CD4+ cell count data and >95% with CD4+ cell count data) to predict risk of delay in linkage to care. In addition, we also identified top high-risk zip codes of delayed linkage. The findings helped public health teams identify high-risk communities of delayed HIV care continuum across Mecklenburg County. The methodology framework can be applied to other regions with HIV epidemic and challenge of delayed linkage to care.

A Multimodality Machine Learning Approach to Differentiate Severe and Nonsevere COVID-19: Model Development and Validation.

BackgroundEffectively and efficiently diagnosing patients who have COVID-19 with the accurate clinical type of the disease is essential to achieve optimal outcomes for the patients as well as to reduce the risk of overloading the health care system. Currently, severe and nonsevere COVID-19 types are differentiated by only a few features, which do not comprehensively characterize the complicated pathological, physiological, and immunological responses to SARS-CoV-2 infection in the different disease types. In addition, these type-defining features may not be readily testable at the time of diagnosis.ObjectiveIn this study, we aimed to use a machine learning approach to understand COVID-19 more comprehensively, accurately differentiate severe and nonsevere COVID-19 clinical types based on multiple medical features, and provide reliable predictions of the clinical type of the disease.MethodsFor this study, we recruited 214 confirmed patients with nonsevere COVID-19 and 148 patients with severe COVID-19. The clinical characteristics (26 features) and laboratory test results (26 features) upon admission were acquired as two input modalities. Exploratory analyses demonstrated that these features differed substantially between two clinical types. Machine learning random forest models based on all the features in each modality as well as on the top 5 features in each modality combined were developed and validated to differentiate COVID-19 clinical types.ResultsUsing clinical and laboratory results independently as input, the random forest models achieved >90% and >95% predictive accuracy, respectively. The importance scores of the input features were further evaluated, and the top 5 features from each modality were identified (age, hypertension, cardiovascular disease, gender, and diabetes for the clinical features modality, and dimerized plasmin fragment D, high sensitivity troponin I, absolute neutrophil count, interleukin 6, and lactate dehydrogenase for the laboratory testing modality, in descending order). Using these top 10 multimodal features as the only input instead of all 52 features combined, the random forest model was able to achieve 97% predictive accuracy.ConclusionsOur findings shed light on how the human body reacts to SARS-CoV-2 infection as a unit and provide insights on effectively evaluating the disease severity of patients with COVID-19 based on more common medical features when gold standard features are not available. We suggest that clinical information can be used as an initial screening tool for self-evaluation and triage, while laboratory test results should be applied when accuracy is the priority.

A novel microfluidic-based assay to determine high-risk pancreatic cysts for surgery utilizing pancreatic cyst fluid: an international multicenter study

Presenter: Ajay Maker MD | University of California San Francisco Background: Pancreatic cysts are increasingly diagnosed in the general population. We previously reported that the majority of cysts resected in the country are low risk lesions on final pathology, despite multiple established clinical guidelines. Thus, selecting which patients harbor a cyst at high-risk of occult malignancy or malignant degeneration for operative resection remains a clinical dilemma. The aim of the current study is to develop a high-throughout molecular diagnostic platform to identify pancreatic cysts at high risk of malignancy or high-grade dysplasia. Methods: Pancreatic cyst fluid from an international multi-center patient cohort were analyzed on a newly constructed microfluidic TaqMan low density array that was customized with our gene target panel of five miRNAs and seven mRNAs. Signatures were validated utilizing a machine learning random forest model for the optimal accuracy to determine high-risk cysts (high-grade dysplasia or invasive cancer) from low risk cysts for operative intervention. Results: Of 358 pancreatic cyst fluids analyzed, 157 cysts were intraductal papillary mucinous neoplasms (IPMNs), 62 were mucinous cystic neoplasms (MCNs), and the remainder were of various non-mucinous histologies. Of the 219 mucinous cysts, 60 (27%) were found to harbor high-grade dysplasia (HGD) or invasive cancer on final pathological examination. The models were established and then internally validated revealing a maximal accuracy of 87% to determine benign from potentially malignant cysts with an out-of-bag error of 8-12%. Further, the assay performed with an accuracy of 81% to identify high-risk from low-risk mucinous lesions (IPMN+MCN) for operative intervention. Conclusion: A high-throughput, cost-effective microfluidics-based molecular diagnostic assay has been developed that identifies high-risk cystic lesions of the pancreas for surgical evaluation. In this large, international, multi-center study, robust validation sets were performed. Performance of the biosignatures are currently being enhanced with clinical and radiographic features.

Machine learning model demonstrates stunting at birth and systemic inflammatory biomarkers as predictors of subsequent infant growth \u2013 a four-year prospective study

BackgroundStunting affects up to one-third of the children in low-to-middle income countries (LMICs) and has been correlated with decline in cognitive capacity and vaccine immunogenicity. Early identification of infants at risk is critical for early intervention and prevention of morbidity. The aim of this study was to investigate patterns of growth in infants up through 48 months of age to assess whether the growth of infants with stunting eventually improved as well as the potential predictors of growth.MethodsHeight-for-age z-scores (HAZ) of children from Matiari (rural site, Pakistan) at birth, 18 months, and 48 months were obtained. Results of serum-based biomarkers collected at 6 and 9 months were recorded. A descriptive analysis of the population was followed by assessment of growth predictors via traditional machine learning random forest models.ResultsOf the 107 children who were followed up till 48 months of age, 51% were stunted (HAZ < − 2) at birth which increased to 54% by 48 months of age. Stunting status for the majority of children at 48 months was found to be the same as at 18 months. Most children with large gains started off stunted or severely stunted, while all of those with notably large losses were not stunted at birth. Random forest models identified HAZ at birth as the most important feature in predicting HAZ at 18 months. Of the biomarkers, AGP (Alpha- 1-acid Glycoprotein), CRP (C-Reactive Protein), and IL1 (interleukin-1) were identified as strong subsequent growth predictors across both the classification and regressor models.ConclusionWe demonstrated that children most children with stunting at birth remained stunted at 48 months of age. Value was added for predicting growth outcomes with the use of traditional machine learning random forest models. HAZ at birth was found to be a strong predictor of subsequent growth in infants up through 48 months of age. Biomarkers of systemic inflammation, AGP, CRP, IL1, were also strong predictors of growth outcomes. These findings provide support for continued focus on interventions prenatally, at birth, and early infancy in children at risk for stunting who live in resource-constrained regions of the world.

Deep profiling of apoptotic pathways with mass cytometry identifies a synergistic drug combination for killing myeloma cells.

Multiple myeloma is an incurable and fatal cancer of immunoglobulin-secreting plasma cells. Most conventional therapies aim to induce apoptosis in myeloma cells but resistance to these drugs often arises and drives relapse. In this study, we sought to identify the best adjunct targets to kill myeloma cells resistant to conventional therapies using deep profiling by mass cytometry (CyTOF). We validated probes to simultaneously detect 26 regulators of cell death, mitosis, cell signaling, and cancer-related pathways at the single-cell level following treatment of myeloma cells with dexamethasone or bortezomib. Time-resolved visualization algorithms and machine learning random forest models (RFMs) delineated putative cell death trajectories and a hierarchy of parameters that specified myeloma cell survival versus apoptosis following treatment. Among these parameters, increased amounts of phosphorylated cAMP response element-binding protein (CREB) and the pro-survival protein, MCL-1, were defining features of cells surviving drug treatment. Importantly, the RFM prediction that the combination of an MCL-1 inhibitor with dexamethasone would elicit potent, synergistic killing of myeloma cells was validated in other cell lines, in vivo preclinical models and primary myeloma samples from patients. Furthermore, CyTOF analysis of patient bone marrow cells clearly identified myeloma cells and their key cell survival features. This study demonstrates the utility of CyTOF profiling at the single-cell level to identify clinically relevant drug combinations and tracking of patient responses for future clinical trials.

On the interpretability of machine learning-based model for predicting hypertension

BackgroundAlthough complex machine learning models are commonly outperforming the traditional simple interpretable models, clinicians find it hard to understand and trust these complex models due to the lack of intuition and explanation of their predictions. The aim of this study to demonstrate the utility of various model-agnostic explanation techniques of machine learning models with a case study for analyzing the outcomes of the machine learning random forest model for predicting the individuals at risk of developing hypertension based on cardiorespiratory fitness data.MethodsThe dataset used in this study contains information of 23,095 patients who underwent clinician-referred exercise treadmill stress testing at Henry Ford Health Systems between 1991 and 2009 and had a complete 10-year follow-up. Five global interpretability techniques (Feature Importance, Partial Dependence Plot, Individual Conditional Expectation, Feature Interaction, Global Surrogate Models) and two local interpretability techniques (Local Surrogate Models, Shapley Value) have been applied to present the role of the interpretability techniques on assisting the clinical staff to get better understanding and more trust of the outcomes of the machine learning-based predictions.ResultsSeveral experiments have been conducted and reported. The results show that different interpretability techniques can shed light on different insights on the model behavior where global interpretations can enable clinicians to understand the entire conditional distribution modeled by the trained response function. In contrast, local interpretations promote the understanding of small parts of the conditional distribution for specific instances.ConclusionsVarious interpretability techniques can vary in their explanations for the behavior of the machine learning model. The global interpretability techniques have the advantage that it can generalize over the entire population while local interpretability techniques focus on giving explanations at the level of instances. Both methods can be equally valid depending on the application need. Both methods are effective methods for assisting clinicians on the medical decision process, however, the clinicians will always remain to hold the final say on accepting or rejecting the outcome of the machine learning models and their explanations based on their domain expertise.

A Phenotype-Based Approach for the Substrate Water Status Forecast of Greenhouse Netted Muskmelon.

Cultivation substrate water status is of great importance to the production of netted muskmelon (Cucumis melo L. var. reticulatus Naud.). A prediction model for the substrate water status would be beneficial in irrigation schedule guidance. In this study, the machine learning random forest model was used to forecast plant substrate water status given the phenotypic traits throughout the muskmelon growing season. Here, two varieties of netted muskmelon, “Wanglu” and “Arus”, were planted in a greenhouse under four substrate water treatments and their phenotypic traits were measured by taking the images within the visible and near-infrared spectrums, respectively. Results showed that a simplified model outperformed the original model in forecasting speed, while it only uses the top five most significant contribution traits. The forecast accuracy reached up to 77.60%, 94.37%, and 90.01% for seedling, vine elongation, and fruit growth stages, respectively. Combining the imaging phenotypic traits and machine learning technique would provide a robust forecast of water status around the plant root zones.