Dataset Information

Study becomes insight: Ecological learning from machine learning.

ABSTRACT: The ecological and environmental science communities have embraced machine learning (ML) for empirical modelling and prediction. However, going beyond prediction to draw insights into underlying functional relationships between response variables and environmental 'drivers' is less straightforward. Deriving ecological insights from fitted ML models requires techniques to extract the 'learning' hidden in the ML models.We revisit the theoretical background and effectiveness of four approaches for deriving insights from ML: ranking independent variable importance (Gini importance, GI; permutation importance, PI; split importance, SI; and conditional permutation importance, CPI), and two approaches for inference of bivariate functional relationships (partial dependence plots, PDP; and accumulated local effect plots, ALE). We also explore the use of a surrogate model for visualization and interpretation of complex multi-variate relationships between response variables and environmental drivers. We examine the challenges and opportunities for extracting ecological insights with these interpretation approaches. Specifically, we aim to improve interpretation of ML models by investigating how effectiveness relates to (a) interpretation algorithm, (b) sample size and (c) the presence of spurious explanatory variables.We base the analysis on simulations with known underlying functional relationships between response and predictor variables, with added white noise and the presence of correlated but non-influential variables. The results indicate that deriving ecological insight is strongly affected by interpretation algorithm and spurious variables, and moderately impacted by sample size. Removing spurious variables improves interpretation of ML models. Meanwhile, increasing sample size has limited value in the presence of spurious variables, but increasing sample size does improves performance once spurious variables are omitted. Among the four ranking methods, SI is slightly more effective than the other methods in the presence of spurious variables, while GI and SI yield higher accuracy when spurious variables are removed. PDP is more effective in retrieving underlying functional relationships than ALE, but its reliability declines sharply in the presence of spurious variables. Visualization and interpretation of the interactive effects of predictors and the response variable can be enhanced using surrogate models, including three-dimensional visualizations and use of loess planes to represent independent variable effects and interactions.Machine learning analysts should be aware that including correlated independent variables in ML models with no clear causal relationship to response variables can interfere with ecological inference. When ecological inference is important, ML models should be constructed with independent variables that have clear causal effects on response variables. While interpreting ML models for ecological inference remains challenging, we show that careful choice of interpretation methods, exclusion of spurious variables and adequate sample size can provide more and better opportunities to 'learn from machine learning'.

SUBMITTER: Yu Q

PROVIDER: S-EPMC9292299 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:Background:Hospital readmission prediction in pediatric hospitals has received little attention. Studies have focused on the readmission frequency analysis stratified by disease and demographic/geographic characteristics but there are no predictive modeling approaches, which may be useful to identify preventable readmissions that constitute a major portion of the cost attributed to readmissions. Objective:To assess the all-cause readmission predictive performance achieved by machine learning techniques in the emergency department of a pediatric hospital in Santiago, Chile. Materials:An all-cause admissions dataset has been collected along six consecutive years in a pediatric hospital in Santiago, Chile. The variables collected are the same used for the determination of the child's treatment administrative cost. Methods:Retrospective predictive analysis of 30-day readmission was formulated as a binary classification problem. We report classification results achieved with various model building approaches after data curation and preprocessing for correction of class imbalance. We compute repeated cross-validation (RCV) with decreasing number of folders to assess performance and sensitivity to effect of imbalance in the test set and training set size. Results:Increase in recall due to SMOTE class imbalance correction is large and statistically significant. The Naive Bayes (NB) approach achieves the best AUC (0.65); however the shallow multilayer perceptron has the best PPV and f-score (5.6 and 10.2, resp.). The NB and support vector machines (SVM) give comparable results if we consider AUC, PPV, and f-score ranking for all RCV experiments. High recall of deep multilayer perceptron is due to high false positive ratio. There is no detectable effect of the number of folds in the RCV on the predictive performance of the algorithms. Conclusions:We recommend the use of Naive Bayes (NB) with Gaussian distribution model as the most robust modeling approach for pediatric readmission prediction, achieving the best results across all training dataset sizes. The results show that the approach could be applied to detect preventable readmissions.

Project description:BackgroundTwitter presents a valuable and relevant social media platform to study the prevalence of information and sentiment on vaping that may be useful for public health surveillance. Machine learning classifiers that identify vaping-relevant tweets and characterize sentiments in them can underpin a Twitter-based vaping surveillance system. Compared with traditional machine learning classifiers that are reliant on annotations that are expensive to obtain, deep learning classifiers offer the advantage of requiring fewer annotated tweets by leveraging the large numbers of readily available unannotated tweets.ObjectiveThis study aims to derive and evaluate traditional and deep learning classifiers that can identify tweets relevant to vaping, tweets of a commercial nature, and tweets with provape sentiments.MethodsWe continuously collected tweets that matched vaping-related keywords over 2 months from August 2018 to October 2018. From this data set of tweets, a set of 4000 tweets was selected, and each tweet was manually annotated for relevance (vape relevant or not), commercial nature (commercial or not), and sentiment (provape or not). Using the annotated data, we derived traditional classifiers that included logistic regression, random forest, linear support vector machine, and multinomial naive Bayes. In addition, using the annotated data set and a larger unannotated data set of tweets, we derived deep learning classifiers that included a convolutional neural network (CNN), long short-term memory (LSTM) network, LSTM-CNN network, and bidirectional LSTM (BiLSTM) network. The unannotated tweet data were used to derive word vectors that deep learning classifiers can leverage to improve performance.ResultsLSTM-CNN performed the best with the highest area under the receiver operating characteristic curve (AUC) of 0.96 (95% CI 0.93-0.98) for relevance, all deep learning classifiers including LSTM-CNN performed better than the traditional classifiers with an AUC of 0.99 (95% CI 0.98-0.99) for distinguishing commercial from noncommercial tweets, and BiLSTM performed the best with an AUC of 0.83 (95% CI 0.78-0.89) for provape sentiment. Overall, LSTM-CNN performed the best across all 3 classification tasks.ConclusionsWe derived and evaluated traditional machine learning and deep learning classifiers to identify vaping-related relevant, commercial, and provape tweets. Overall, deep learning classifiers such as LSTM-CNN had superior performance and had the added advantage of requiring no preprocessing. The performance of these classifiers supports the development of a vaping surveillance system.

Project description:Networks of trophic links (food webs) are used to describe and understand mechanistic routes for translocation of energy (biomass) between species. However, a relatively low proportion of ecosystems have been studied using food web approaches due to difficulties in making observations on large numbers of species. In this paper we demonstrate that Machine Learning of food webs, using a logic-based approach called A/ILP, can generate plausible and testable food webs from field sample data. Our example data come from a national-scale Vortis suction sampling of invertebrates from arable fields in Great Britain. We found that 45 invertebrate species or taxa, representing approximately 25% of the sample and about 74% of the invertebrate individuals included in the learning, were hypothesized to be linked. As might be expected, detritivore Collembola were consistently the most important prey. Generalist and omnivorous carabid beetles were hypothesized to be the dominant predators of the system. We were, however, surprised by the importance of carabid larvae suggested by the machine learning as predators of a wide variety of prey. High probability links were hypothesized for widespread, potentially destabilizing, intra-guild predation; predictions that could be experimentally tested. Many of the high probability links in the model have already been observed or suggested for this system, supporting our contention that A/ILP learning can produce plausible food webs from sample data, independent of our preconceptions about "who eats whom." Well-characterised links in the literature correspond with links ascribed with high probability through A/ILP. We believe that this very general Machine Learning approach has great power and could be used to extend and test our current theories of agricultural ecosystem dynamics and function. In particular, we believe it could be used to support the development of a wider theory of ecosystem responses to environmental change.

Project description:BACKGROUND:Although geriatric depression is prevalent, diagnosis using self-reporting instruments has limitations when measuring the depressed mood of older adults in a community setting. Ecological momentary assessment (EMA) by using wearable devices could be used to collect data to classify older adults into depression groups. OBJECTIVE:The objective of this study was to develop a machine learning algorithm to predict the classification of depression groups among older adults living alone. We focused on utilizing diverse data collected through a survey, an Actiwatch, and an EMA report related to depression. METHODS:The prediction model using machine learning was developed in 4 steps: (1) data collection, (2) data processing and representation, (3) data modeling (feature engineering and selection), and (4) training and validation to test the prediction model. Older adults (N=47), living alone in community settings, completed an EMA to report depressed moods 4 times a day for 2 weeks between May 2017 and January 2018. Participants wore an Actiwatch that measured their activity and ambient light exposure every 30 seconds for 2 weeks. At baseline and the end of the 2-week observation, depressive symptoms were assessed using the Korean versions of the Short Geriatric Depression Scale (SGDS-K) and the Hamilton Depression Rating Scale (K-HDRS). Conventional classification based on binary logistic regression was built and compared with 4 machine learning models (the logit, decision tree, boosted trees, and random forest models). RESULTS:On the basis of the SGDS-K and K-HDRS, 38% (18/47) of the participants were classified into the probable depression group. They reported significantly lower scores of normal mood and physical activity and higher levels of white and red, green, and blue (RGB) light exposures at different degrees of various 4-hour time frames (all P<.05). Sleep efficiency was chosen for modeling through feature selection. Comparing diverse combinations of the selected variables, daily mean EMA score, daily mean activity level, white and RGB light at 4:00 pm to 8:00 pm exposure, and daily sleep efficiency were selected for modeling. Conventional classification based on binary logistic regression had a good model fit (accuracy: 0.705; precision: 0.770; specificity: 0.859; and area under receiver operating characteristic curve or AUC: 0.754). Among the 4 machine learning models, the logit model had the best fit compared with the others (accuracy: 0.910; precision: 0.929; specificity: 0.940; and AUC: 0.960). CONCLUSIONS:This study provides preliminary evidence for developing a machine learning program to predict the classification of depression groups in older adults living alone. Clinicians should consider using this method to identify underdiagnosed subgroups and monitor daily progression regarding treatment or therapeutic intervention in the community setting. Furthermore, more efforts are needed for researchers and clinicians to diversify data collection methods by using a survey, EMA, and a sensor.

Dataset Information

Study becomes insight: Ecological learning from machine learning.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets