ResearchPad - support-vector-machines https://www.researchpad.co Default RSS Feed en-us © 2020 Newgen KnowledgeWorks <![CDATA[Deep learning assisted detection of glaucomatous optic neuropathy and potential designs for a generalizable model]]> https://www.researchpad.co/article/elastic_article_14620 To evaluate ways to improve the generalizability of a deep learning algorithm for identifying glaucomatous optic neuropathy (GON) using a limited number of fundus photographs, as well as the key features being used for classification.MethodsA total of 944 fundus images from Taipei Veterans General Hospital (TVGH) were retrospectively collected. Clinical and demographic characteristics, including structural and functional measurements of the images with GON, were recorded. Transfer learning based on VGGNet was used to construct a convolutional neural network (CNN) to identify GON. To avoid missing cases with advanced GON, an ensemble model was adopted in which a support vector machine classifier would make final classification based on cup-to-disc ratio if the CNN classifier had low-confidence score. The CNN classifier was first established using TVGH dataset, and then fine-tuned by combining the training images of TVGH and Drishti-GS datasets. Class activation map (CAM) was used to identify key features used for CNN classification. Performance of each classifier was determined through area under receiver operating characteristic curve (AUC) and compared with the ensemble model by diagnostic accuracy.ResultsIn 187 TVGH test images, the accuracy, sensitivity, and specificity of the CNN classifier were 95.0%, 95.7%, and 94.2%, respectively, and the AUC was 0.992 compared to the 92.8% accuracy rate of the ensemble model. For the Drishti-GS test images, the accuracy of the CNN, the fine-tuned CNN and ensemble model was 33.3%, 80.3%, and 80.3%, respectively. The CNN classifier did not misclassify images with moderate to severe diseases. Class-discriminative regions revealed by CAM co-localized with known characteristics of GON.ConclusionsThe ensemble model or a fine-tuned CNN classifier may be potential designs to build a generalizable deep learning model for glaucoma detection when large image databases are not available. ]]> <![CDATA[ECG-based prediction algorithm for imminent malignant ventricular arrhythmias using decision tree]]> https://www.researchpad.co/article/elastic_article_14548 Spontaneous prediction of malignant ventricular arrhythmia (MVA) is useful to avoid delay in rescue operations. Recently, researchers have developed several algorithms to predict MVA using various features derived from electrocardiogram (ECG). However, there are several unresolved issues regarding MVA prediction such as the effect of number of ECG features on a prediction remaining unclear, possibility that an alert for occurring MVA may arrive very late and uncertainty in the performance of the algorithm predicting MVA minutes before onset. To overcome the aforementioned problems, this research conducts an in-depth study on the number and types of ECG features that are implemented in a decision tree classifier. In addition, this research also investigates an algorithm’s execution time before the occurrence of MVA to minimize delays in warnings for MVA. Lastly, this research aims to study both the sensitivity and specificity of an algorithm to reveal the performance of MVA prediction algorithms from time to time. To strengthen the results of analysis, several classifiers such as support vector machine and naive Bayes are also examined for the purpose of comparison study. There are three phases required to achieve the objectives. The first phase is literature review on existing relevant studies. The second phase deals with design and development of four modules for predicting MVA. Rigorous experiments are performed in the feature selection and classification modules. The results show that eight ECG features with decision tree classifier achieved good prediction performance in terms of execution time and sensitivity. In addition, the results show that the highest percentage for sensitivity and specificity is 95% and 90% respectively, in the fourth 5-minute interval (15.1 minutes–20 minutes) that preceded the onset of an arrhythmia event. Such results imply that the fourth 5-minute interval would be the best time to perform prediction.

]]>
<![CDATA[Insight into the protein solubility driving forces with neural attention]]> https://www.researchpad.co/article/elastic_article_13832 The solubility of proteins is a crucial biophysical aspect when it comes to understanding many human diseases and to improve the industrial processes for protein production. Due to its relevance, computational methods have been devised in order to study and possibly optimize the solubility of proteins. In this work we apply a deep-learning technique, called neural attention to predict protein solubility while “opening” the model itself to interpretability, even though Machine Learning models are usually considered black boxes. Thank to the attention mechanism, we show that i) our model implicitly learns complex patterns related to emergent, protein folding-related, aspects such as to recognize β-amyloidosis regions and that ii) the N-and C-termini are the regions with the highes signal fro solubility prediction. When it comes to enhancing the solubility of proteins, we, for the first time, propose to investigate the synergistic effects of tandem mutations instead of “single” mutations, suggesting that this could minimize the number of required proposed mutations.

]]>
<![CDATA[Forecasting the monthly incidence rate of brucellosis in west of Iran using time series and data mining from 2010 to 2019]]> https://www.researchpad.co/article/elastic_article_13811 The identification of statistical models for the accurate forecast and timely determination of the outbreak of infectious diseases is very important for the healthcare system. Thus, this study was conducted to assess and compare the performance of four machine-learning methods in modeling and forecasting brucellosis time series data based on climatic parameters.MethodsIn this cohort study, human brucellosis cases and climatic parameters were analyzed on a monthly basis for the Qazvin province–located in northwestern Iran- over a period of 9 years (2010–2018). The data were classified into two subsets of education (80%) and testing (20%). Artificial neural network methods (radial basis function and multilayer perceptron), support vector machine and random forest were fitted to each set. Performance analysis of the models were done using the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Root Error (MARE), and R2 criteria.ResultsThe incidence rate of the brucellosis in Qazvin province was 27.43 per 100,000 during 2010–2019. Based on our results, the values of the RMSE (0.22), MAE (0.175), MARE (0.007) criteria were smaller for the multilayer perceptron neural network than their values in the other three models. Moreover, the R2 (0.99) value was bigger in this model. Therefore, the multilayer perceptron neural network exhibited better performance in forecasting the studied data. The average wind speed and mean temperature were the most effective climatic parameters in the incidence of this disease.ConclusionsThe multilayer perceptron neural network can be used as an effective method in detecting the behavioral trend of brucellosis over time. Nevertheless, further studies focusing on the application and comparison of these methods are needed to detect the most appropriate forecast method for this disease. ]]> <![CDATA[A compound attributes-based predictive model for drug induced liver injury in humans]]> https://www.researchpad.co/article/Ndeb57c49-a1cc-41d4-9618-08dc56c45dac

Drug induced liver injury (DILI) is one of the key safety concerns in drug development. To assess the likelihood of drug candidates with potential adverse reactions of liver, we propose a compound attributes-based approach to predicting hepatobiliary disorders that are routinely reported to US Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS). Specifically, we developed a support vector machine (SVM) model with recursive feature extraction, based on physicochemical and structural properties of compounds as model input. Cross validation demonstrates that the predictive model has a robust performance with averaged 70% of both sensitivity and specificity over 500 trials. An independent validation was performed on public benchmark drugs and the results suggest potential utility of our model for identifying safety alerts. This in silico approach, upon further validation, would ultimately be implemented, together with other in vitro safety assays, for screening compounds early in drug development.

]]>
<![CDATA[Neuroimaging modality fusion in Alzheimer’s classification using convolutional neural networks]]> https://www.researchpad.co/article/N4bce0426-e39d-45a0-9dc9-42db4f6cba04

Automated methods for Alzheimer’s disease (AD) classification have the potential for great clinical benefits and may provide insight for combating the disease. Machine learning, and more specifically deep neural networks, have been shown to have great efficacy in this domain. These algorithms often use neurological imaging data such as MRI and FDG PET, but a comprehensive and balanced comparison of the MRI and amyloid PET modalities has not been performed. In order to accurately determine the relative strength of each imaging variant, this work performs a comparison study in the context of Alzheimer’s dementia classification using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset with identical neural network architectures. Furthermore, this work analyzes the benefits of using both modalities in a fusion setting and discusses how these data types may be leveraged in future AD studies using deep learning.

]]>
<![CDATA[Adaptive multi-degree of freedom Brain Computer Interface using online feedback: Towards novel methods and metrics of mutual adaptation between humans and machines for BCI]]> https://www.researchpad.co/article/5c89771ad5eed0c4847d2469

This paper proposes a novel adaptive online-feedback methodology for Brain Computer Interfaces (BCI). The method uses ElectroEncephaloGraphic (EEG) signals and combines motor with speech imagery to allow for tasks that involve multiple degrees of freedom (DoF). The main approach utilizes the covariance matrix descriptor as feature, and the Relevance Vector Machines (RVM) classifier. The novel contributions include, (1) a new method to select representative data to update the RVM model, and (2) an online classifier which is an adaptively-weighted mixture of RVM models to account for the users’ exploration and exploitation processes during the learning phase. Instead of evaluating the subjects’ performance solely based on the conventional metric of accuracy, we analyze their skill’s improvement based on 3 other criteria, namely the confusion matrix’s quality, the separability of the data, and their instability. After collecting calibration data for 8 minutes in the first run, 8 participants were able to control the system while receiving visual feedback in the subsequent runs. We observed significant improvement in all subjects, including two of them who fell into the BCI illiteracy category. Our proposed BCI system complements the existing approaches in several aspects. First, the co-adaptation paradigm not only adapts the classifiers, but also allows the users to actively discover their own way to use the BCI through their exploration and exploitation processes. Furthermore, the auto-calibrating system can be used immediately with a minimal calibration time. Finally, this is the first work to combine motor and speech imagery in an online feedback experiment to provide multiple DoF for BCI control applications.

]]>
<![CDATA[Structure and variability of delay activity in premotor cortex]]> https://www.researchpad.co/article/5c990204d5eed0c484b9749c

Voluntary movements are widely considered to be planned before they are executed. Recent studies have hypothesized that neural activity in motor cortex during preparation acts as an ‘initial condition’ which seeds the proceeding neural dynamics. Here, we studied these initial conditions in detail by investigating 1) the organization of neural states for different reaches and 2) the variance of these neural states from trial to trial. We examined population-level responses in macaque premotor cortex (PMd) during the preparatory stage of an instructed-delay center-out reaching task with dense target configurations. We found that after target onset the neural activity on single trials converges to neural states that have a clear low-dimensional structure which is organized by both the reach endpoint and maximum speed of the following reach. Further, we found that variability of the neural states during preparation resembles the spatial variability of reaches made in the absence of visual feedback: there is less variability in direction than distance in neural state space. We also used offline decoding to understand the implications of this neural population structure for brain-machine interfaces (BMIs). We found that decoding of angle between reaches is dependent on reach distance, while decoding of arc-length is independent. Thus, it might be more appropriate to quantify decoding performance for discrete BMIs by using arc-length between reach end-points rather than the angle between them. Lastly, we show that in contrast to the common notion that direction can better be decoded than distance, their decoding capabilities are comparable. These results provide new insights into the dynamical neural processes that underline motor control and can inform the design of BMIs.

]]>
<![CDATA[Selection of the optimal trading model for stock investment in different industries]]> https://www.researchpad.co/article/5c6dc9d9d5eed0c48452a2e0

In general, the stock prices of the same industry have a similar trend, but those of different industries do not. When investing in stocks of different industries, one should select the optimal model from lots of trading models for each industry because any model may not be suitable for capturing the stock trends of all industries. However, the study has not been carried out at present. In this paper, firstly we select 424 S&P 500 index component stocks (SPICS) and 185 CSI 300 index component stocks (CSICS) as the research objects from 2010 to 2017, divide them into 9 industries such as finance and energy respectively. Secondly, we apply 12 widely used machine learning algorithms to generate stock trading signals in different industries and execute the back-testing based on the trading signals. Thirdly, we use a non-parametric statistical test to evaluate whether there are significant differences among the trading performance evaluation indicators (PEI) of different models in the same industry. Finally, we propose a series of rules to select the optimal models for stock investment of every industry. The analytical results on SPICS and CSICS show that we can find the optimal trading models for each industry based on the statistical tests and the rules. Most importantly, the PEI of the best algorithms can be significantly better than that of the benchmark index and “Buy and Hold” strategy. Therefore, the algorithms can be used for making profits from industry stock trading.

]]>
<![CDATA[Plant leaf tooth feature extraction]]> https://www.researchpad.co/article/5c6dc9b6d5eed0c48452a077

Leaf tooth can indicate several systematically informative features and is extremely useful for circumscribing fossil leaf taxa. Moreover, it can help discriminate species or even higher taxa accurately. Previous studies extract features that are not strictly defined in botany; therefore, a uniform standard to compare the accuracies of various feature extraction methods cannot be used. For efficient and automatic retrieval of plant leaves from a leaf database, in this study, we propose an image-based description and measurement of leaf teeth by referring to the leaf structure classification system in botany. First, image preprocessing is carried out to obtain a binary map of plant leaves. Then, corner detection based on the curvature scale-space (CSS) algorithm is used to extract the inflection point from the edges; next, the leaf tooth apex is extracted by screening the convex points; then, according to the definition of the leaf structure, the characteristics of the leaf teeth are described and measured in terms of number of orders of teeth, tooth spacing, number of teeth, sinus shape, and tooth shape. In this manner, data extracted from the algorithm can not only be used to classify plants, but also provide scientific and standardized data to understand the history of plant evolution. Finally, to verify the effectiveness of the extraction method, we used simple linear discriminant analysis and multiclass support vector machine to classify leaves. The results show that the proposed method achieves high accuracy that is superior to that of other methods.

]]>
<![CDATA[Resolution invariant wavelet features of melanoma studied by SVM classifiers]]> https://www.researchpad.co/article/5c648cd2d5eed0c484c81893

This article refers to the Computer Aided Diagnosis of the melanoma skin cancer. We derive wavelet-based features of melanoma from the dermoscopic images of pigmental skin lesions and apply binary C-SVM classifiers to discriminate malignant melanoma from dysplastic nevus. The aim of this research is to select the most efficient model of the SVM classifier for various image resolutions and to search for the best resolution-invariant wavelet bases. We show AUC as a function of the wavelet number and SVM kernels optimized by the Bayesian search for two independent data sets. Our results are compatible with the previous experiments to discriminate melanoma in dermoscopy images with ensembling and feed-forward neural networks.

]]>
<![CDATA[Label-free classification of cells based on supervised machine learning of subcellular structures]]> https://www.researchpad.co/article/5c59fec3d5eed0c48413541b

It is demonstrated that cells can be classified by pattern recognition of the subcellular structure of non-stained live cells, and the pattern recognition was performed by machine learning. Human white blood cells and five types of cancer cell lines were imaged by quantitative phase microscopy, which provides morphological information without staining quantitatively in terms of optical thickness of cells. Subcellular features were then extracted from the obtained images as training data sets for the machine learning. The built classifier successfully classified WBCs from cell lines (area under ROC curve = 0.996). This label-free, non-cytotoxic cell classification based on the subcellular structure of QPM images has the potential to serve as an automated diagnosis of single cells.

]]>
<![CDATA[Using supervised learning to select audit targets in performance-based financing in health: An example from Zambia]]> https://www.researchpad.co/article/5c59ff00d5eed0c4841358af

Independent verification is a critical component of performance-based financing (PBF) in health care, in which facilities are offered incentives to increase the volume of specific services but the same incentives may lead them to over-report. We examine alternative strategies for targeted sampling of health clinics for independent verification. Specifically, we empirically compare several methods of random sampling and predictive modeling on data from a Zambian PBF pilot that contains reported and verified performance for quantity indicators of 140 clinics. Our results indicate that machine learning methods, particularly Random Forest, outperform other approaches and can increase the cost-effectiveness of verification activities.

]]>
<![CDATA[Software-aided workflow for predicting protease-specific cleavage sites using physicochemical properties of the natural and unnatural amino acids in peptide-based drug discovery]]> https://www.researchpad.co/article/5c3e50a9d5eed0c484d840db

Peptide drugs have been used in the treatment of multiple pathologies. During peptide discovery, it is crucially important to be able to map the potential sites of cleavages of the proteases. This knowledge is used to later chemically modify the peptide drug to adapt it for the therapeutic use, making peptide stable against individual proteases or in complex medias. In some other cases it needed to make it specifically unstable for some proteases, as peptides could be used as a system to target delivery drugs on specific tissues or cells. The information about proteases, their sites of cleavages and substrates are widely spread across publications and collected in databases such as MEROPS. Therefore, it is possible to develop models to improve the understanding of the potential peptide drug proteolysis. We propose a new workflow to derive protease specificity rules and predict the potential scissile bonds in peptides for individual proteases. WebMetabase stores the information from experimental or external sources in a chemically aware database where each peptide and site of cleavage is represented as a sequence of structural blocks connected by amide bonds and characterized by its physicochemical properties described by Volsurf descriptors. Thus, this methodology could be applied in the case of non-standard amino acid. A frequency analysis can be performed in WebMetabase to discover the most frequent cleavage sites. These results were used to train several models using logistic regression, support vector machine and ensemble tree classifiers to map cleavage sites for several human proteases from four different families (serine, cysteine, aspartic and matrix metalloproteases). Finally, we compared the predictive performance of the developed models with other available public tools PROSPERous and SitePrediction.

]]>
<![CDATA[Poincaré plot analysis of cerebral blood flow signals: Feature extraction and classification methods for apnea detection]]> https://www.researchpad.co/article/5c141ee1d5eed0c484d28a77

Objective

Rheoencephalography is a simple and inexpensive technique for cerebral blood flow assessment, however, it is not used in clinical practice since its correlation to clinical conditions has not yet been extensively proved. The present study investigates the ability of Poincaré Plot descriptors from rheoencephalography signals to detect apneas in volunteers.

Methods

A group of 16 subjects participated in the study. Rheoencephalography data from baseline and apnea periods were recorded and Poincaré Plot descriptors were extracted from the reconstructed attractors with different time lags (τ). Among the set of extracted features, those presenting significant differences between baseline and apnea recordings were used as inputs to four different classifiers to optimize the apnea detection.

Results

Three features showed significant differences between apnea and baseline signals: the Poincaré Plot ratio (SDratio), its correlation (R) and the Complex Correlation Measure (CCM). Those differences were optimized for time lags smaller than those recommended in previous works for other biomedical signals, all of them being lower than the threshold established by the position of the inflection point in the CCM curves. The classifier showing the best performance was the classification tree, with 81% accuracy and an area under the curve of the receiver operating characteristic of 0.927. This performance was obtained using a single input parameter, either SDratio or R.

Conclusions

Poincaré Plot features extracted from the attractors of rheoencephalographic signals were able to track cerebral blood flow changes provoked by breath holding. Even though further validation with independent datasets is needed, those results suggest that nonlinear analysis of rheoencephalography might be a useful approach to assess the correlation of cerebral impedance with clinical changes.

]]>
<![CDATA[Automatically recognizing strategic cooperative behaviors in various situations of a team sport]]> https://www.researchpad.co/article/5c22a0bfd5eed0c4849ec088

Understanding multi-agent cooperative behavior is challenging in various scientific and engineering domains. In some cases, such as team sports, many cooperative behaviors can be visually categorized and labeled manually by experts. However, these actions which are manually categorized with the same label based on its function have low spatiotemporal similarity. In other words, it is difficult to find similar and different structures of the motions with the same and different labels, respectively. Here, we propose an automatic recognition system for strategic cooperative plays, which are the minimal, basic, and diverse plays in a ball game. Using player’s moving distance, geometric information, and distances among players, the proposed method accurately discriminated not only the cooperative plays in a primary area, i.e., near the ball, but also those distant from a primary area. We also propose a method to classify more detailed types of cooperative plays in various situations. The proposed framework, which sheds light on inconspicuous players to play important roles, could have a potential to detect well-defined and labeled cooperative behaviors.

]]>
<![CDATA[Wavelet analysis of oximetry recordings to assist in the automated detection of moderate-to-severe pediatric sleep apnea-hypopnea syndrome]]> https://www.researchpad.co/article/5c141eb6d5eed0c484d27e15

Background

The gold standard for pediatric sleep apnea hypopnea syndrome (SAHS) is overnight polysomnography, which has several limitations. Thus, simplified diagnosis techniques become necessary.

Objective

The aim of this study is twofold: (i) to analyze the blood oxygen saturation (SpO2) signal from nocturnal oximetry by means of features from the wavelet transform in order to characterize pediatric SAHS; (ii) to evaluate the usefulness of the extracted features to assist in the detection of pediatric SAHS.

Methods

981 SpO2 signals from children ranging 2–13 years of age were used. Discrete wavelet transform (DWT) was employed due to its suitability to deal with non-stationary signals as well as the ability to analyze the SAHS-related low frequency components of the SpO2 signal with high resolution. In addition, 3% oxygen desaturation index (ODI3), statistical moments and power spectral density (PSD) features were computed. Fast correlation-based filter was applied to select a feature subset. This subset fed three classifiers (logistic regression, support vector machines (SVM), and multilayer perceptron) trained to determine the presence of moderate-to-severe pediatric SAHS (apnea-hypopnea index cutoff ≥ 5 events per hour).

Results

The wavelet entropy and features computed in the D9 detail level of the DWT reached significant differences associated with the presence of SAHS. All the proposed classifiers fed with a selected feature subset composed of ODI3, statistical moments, PSD, and DWT features outperformed every single feature. SVM reached the highest performance. It achieved 84.0% accuracy (71.9% sensitivity, 91.1% specificity), outperforming state-of-the-art studies in the detection of moderate-to-severe SAHS using the SpO2 signal alone.

Conclusion

Wavelet analysis could be a reliable tool to analyze the oximetry signal in order to assist in the automated detection of moderate-to-severe pediatric SAHS. Hence, pediatric subjects suffering from moderate-to-severe SAHS could benefit from an accurate simplified screening test only using the SpO2 signal.

]]>
<![CDATA[Improving the calling of non-invasive prenatal testing on 13-/18-/21-trisomy by support vector machine discrimination]]> https://www.researchpad.co/article/5c117b86d5eed0c48469989c

With the advance of next-generation sequencing (NGS) technologies, non-invasive prenatal testing (NIPT) has been developed and employed in fetal aneuploidy screening on 13-/18-/21-trisomies through detecting cell-free fetal DNA (cffDNA) in maternal blood. Although Z-test is widely used in NIPT NGS data analysis, there is still necessity to improve its accuracy for reducing a) false negatives and false positives, and b) the ratio of unclassified data, so as to lower the potential harm to patients as well as the induced cost of retests. Combining the multiple Z-tests with indexes of clinical signs and quality control, features were collected from the known samples and scaled for model training using support vector machine (SVM). We trained SVM models from the qualified NIPT NGS data that Z-test can discriminate and tested the performance on the data that Z-test cannot discriminate. On screenings of 13-/18-/21-trisomies, the trained SVM models achieved 100% accuracies in both internal validations and unknown sample predictions. It is shown that other machine learning (ML) models can also achieve similar high accuracy, and SVM model is most robust in this study. Moreover, four false positives and four false negatives caused by Z-test were corrected by using the SVM models. To our knowledge, this is one of the earliest studies to employ SVM in NIPT NGS data analysis. It is expected to replace Z-test in clinical practice.

]]>
<![CDATA[Estimation of paddy rice leaf area index using machine learning methods based on hyperspectral data from multi-year experiments]]> https://www.researchpad.co/article/5c117b6fd5eed0c484699303

The performance of three machine learning methods (support vector regression, random forests and artificial neural network) for estimating the LAI of paddy rice was evaluated in this study. Traditional univariate regression models involving narrowband NDVI with optimized band combinations as well as linear multivariate calibration partial least squares regression models were also evaluated for comparison. A four year field-collected dataset was used to test the robustness of LAI estimation models against temporal variation. The partial least squares regression and three machine learning methods were built on the raw hyperspectral reflectance and the first derivative separately. Two different rules were used to determine the models’ key parameters. The results showed that the combination of the red edge and NIR bands (766 nm and 830 nm) as well as the combination of SWIR bands (1114 nm and 1190 nm) were optimal for producing the narrowband NDVI. The models built on the first derivative spectra yielded more accurate results than the corresponding models built on the raw spectra. Properly selected model parameters resulted in comparable accuracy and robustness with the empirical optimal parameter and significantly reduced the model complexity. The machine learning methods were more accurate and robust than the VI methods and partial least squares regression. When validating the calibrated models against the standalone validation dataset, the VI method yielded a validation RMSE value of 1.17 for NDVI(766,830) and 1.01 for NDVI(1114,1190), while the best models for the partial least squares, support vector machine and artificial neural network methods yielded validation RMSE values of 0.84, 0.82, 0.67 and 0.84, respectively. The RF models built on the first derivative spectra with mtry = 10 showed the highest potential for estimating the LAI of paddy rice.

]]>
<![CDATA[Characterising risk of in-hospital mortality following cardiac arrest using machine learning: A retrospective international registry study]]> https://www.researchpad.co/article/5c0ae470d5eed0c484589b1f

Background

Resuscitated cardiac arrest is associated with high mortality; however, the ability to estimate risk of adverse outcomes using existing illness severity scores is limited. Using in-hospital data available within the first 24 hours of admission, we aimed to develop more accurate models of risk prediction using both logistic regression (LR) and machine learning (ML) techniques, with a combination of demographic, physiologic, and biochemical information.

Methods and findings

Patient-level data were extracted from the Australian and New Zealand Intensive Care Society (ANZICS) Adult Patient Database for patients who had experienced a cardiac arrest within 24 hours prior to admission to an intensive care unit (ICU) during the period January 2006 to December 2016. The primary outcome was in-hospital mortality. The models were trained and tested on a dataset (split 90:10) including age, lowest and highest physiologic variables during the first 24 hours, and key past medical history. LR and 5 ML approaches (gradient boosting machine [GBM], support vector classifier [SVC], random forest [RF], artificial neural network [ANN], and an ensemble) were compared to the APACHE III and Australian and New Zealand Risk of Death (ANZROD) predictions. In all, 39,566 patients from 186 ICUs were analysed. Mean (±SD) age was 61 ± 17 years; 65% were male. Overall in-hospital mortality was 45.5%. Models were evaluated in the test set. The APACHE III and ANZROD scores demonstrated good discrimination (area under the receiver operating characteristic curve [AUROC] = 0.80 [95% CI 0.79–0.82] and 0.81 [95% CI 0.8–0.82], respectively) and modest calibration (Brier score 0.19 for both), which was slightly improved by LR (AUROC = 0.82 [95% CI 0.81–0.83], DeLong test, p < 0.001). Discrimination was significantly improved using ML models (ensemble and GBM AUROCs = 0.87 [95% CI 0.86–0.88], DeLong test, p < 0.001), with an improvement in performance (Brier score reduction of 22%). Explainability models were created to assist in identifying the physiologic features that most contributed to an individual patient’s survival. Key limitations include the absence of pre-hospital data and absence of external validation.

Conclusions

ML approaches significantly enhance predictive discrimination for mortality following cardiac arrest compared to existing illness severity scores and LR, without the use of pre-hospital data. The discriminative ability of these ML models requires validation in external cohorts to establish generalisability.

]]>