Supervised Classification Algorithms Research Articles

An RBF Neural Network Clustering Algorithm Based on K-Nearest Neighbor

Neural network is a supervised classification algorithm which can deal with high complexity and nonlinear data analysis. Supervised algorithm needs some known labels in the training process, and then corrects parameters through backpropagation method. However, due to the lack of marked labels, existing literature mostly uses Auto-Encoder to reduce the dimension of data when facing of clustering problems. This paper proposes an RBF (Radial Basis Function) neural network clustering algorithm based on K-nearest neighbors theory, which first uses K-means algorithm for preclassification, and then constructs self-supervised labels based on K-nearest neighbors theory for backpropagation. The algorithm in this paper belongs to a self-supervised neural network clustering algorithm, and it also makes the neural network truly have the ability of self-decision-making and self-optimization. From the experimental results of the artificial data sets and the UCI data sets, it can be proved that the proposed algorithm has excellent adaptability and robustness.

Predicting Activity Duration in Smart Sensing Environments Using Synthetic Data and Partial Least Squares Regression: The Case of Dementia Patients.

The accurate recognition of activities is fundamental for following up on the health progress of people with dementia (PwD), thereby supporting subsequent diagnosis and treatments. When monitoring the activities of daily living (ADLs), it is feasible to detect behaviour patterns, parse out the disease evolution, and consequently provide effective and timely assistance. However, this task is affected by uncertainties derived from the differences in smart home configurations and the way in which each person undertakes the ADLs. One adjacent pathway is to train a supervised classification algorithm using large-sized datasets; nonetheless, obtaining real-world data is costly and characterized by a challenging recruiting research process. The resulting activity data is then small and may not capture each person’s intrinsic properties. Simulation approaches have risen as an alternative efficient choice, but synthetic data can be significantly dissimilar compared to real data. Hence, this paper proposes the application of Partial Least Squares Regression (PLSR) to approximate the real activity duration of various ADLs based on synthetic observations. First, the real activity duration of each ADL is initially contrasted with the one derived from an intelligent environment simulator. Following this, different PLSR models were evaluated for estimating real activity duration based on synthetic variables. A case study including eight ADLs was considered to validate the proposed approach. The results revealed that simulated and real observations are significantly different in some ADLs (p-value < 0.05), nevertheless synthetic variables can be further modified to predict the real activity duration with high accuracy ().

Mapping land cover and forest density in Zagros forests of Khuzestan province in Iran: A study based on Sentinel-2, Google Earth and field data

Zagros forests in western Iran have widely been destroyed because of various reasons. This study was performed to provide the land cover and forest density maps in Zagros forests of Khuzestan province using Sentinel-2, Google Earth and field data. The forest boundary in Khuzestan province was digitized in Google Earth. Sentinel-2 satellite images were provided for the study area. One 1:25000 index sheet of Iranian Mapping Organization (IMO) was selected as pilot area in the province. Sentinel-2 image of the pilot area was classified using different supervised classification algorithms to select the best algorithm for land cover mapping in Khuzestan province. In addition, to evaluate the accuracy of Google Earth data, field sampling was performed using random plots in different land covers. Field data of forest plots were applied to investigate tree canopy cover percent (forest density), as well. Classification of Sentinel-2 image in Zagros area of Khuzestan province was done using the best algorithm and the land cover was obtained. The forest density map was also obtained using a linear regression model between tree canopy cover percent (obtained from field plots) and normalized difference vegetation index (NDVI) (obtained from NDVI map). Finally, the accuracy of land cover map was assessed by some square plots on Google Earth. Results demonstrated that support vector machine (SVM) algorithm had the highest accuracy for land cover mapping. Results also showed that Google Earth images had a good accuracy in the Zagros forests of Khuzestan province. Results demonstrated that NDVI has been a good predicator to estimate tree canopy cover in the study area. Based on results, an area of 443,091.22 ha is covered by Zagros forests in Khuzestan province. Results of accuracy assessment of the land cover map showed the good accuracy of this map in Khuzestan province (overall accuracy: 91% and kappa index: 0.83). For optimum management of Zagros forests, it is suggested that the land cover and forest density mapping will be performed using SVM algorithm, NDVI, and Sentinel-2 satellite images in Zagros forests of Khuzestan province in the certain periods.

Identification and Validation of a Five-Gene Diagnostic Signature for Preeclampsia.

Preeclampsia is the leading cause of morbidity and mortality for mothers and newborns worldwide. Despite extensive efforts made to understand the underlying pathology of preeclampsia, there is still no clinically useful effective tool for the early diagnosis of preeclampsia. In this study, we conducted a retrospectively multicenter discover-validation study to develop and validate a novel biomarker for preeclampsia diagnosis. We identified 38 differentially expressed genes (DEGs) involved in preeclampsia in a case-control study by analyzing expression profiles in the discovery cohort. We developed a 5-mRNA signature (termed PE5-signature) to diagnose preeclampsia from 38 DEGs using recursive feature elimination with a random forest supervised classification algorithm, including ENG, KRT80, CEBPA, RDH13 and WASH9P. The PE5-signature showed high accuracy in discriminating preeclampsia from controls with a receiver operating characteristic area under the curve value (AUC) of 0.971, a sensitivity of 0.842 and a specificity of 0.950. The PE5-signature was then validated in an independent case-control study and achieved a reliable and robust predictive performance with an AUC of 0.929, a sensitivity of 0.696, and a specificity of 0.946. In summary, we have developed and validated a five-mRNA biomarker panel as a risk assessment tool to assist in the detection of preeclampsia. This gene panel has potential clinical value for early preeclampsia diagnosis and may help us better understand the precise mechanisms involved.

Classification and simulation of process of linear change for grip force at different grip speeds by using supervised learning based on sEMG

Bionic prosthesis can help the disabled rebuild the athletic ability of their missing limbs. Biomimetic hand controlled by surface electromyography (sEMG) has been widely studied in recent years. However, a lot of researchers focus on the field of movement recognition, ignoring the precise control of grip speed and force in the design of the prosthesis. The study presents a new means to simulate the process of linear changes of grip force by using sEMG, which provides a simple and practicable way of precise control of the speed and strength of force for the myoelectric hand. We design a synchronous acquisition device for sEMG signal and grip force signal. A novel preprocessing methodology for the raw sEMG signal is introduced, which is adaptive filter based on the Normalized Least Mean Square (NLMS). Compared the traditional preprocessing method that uses a notch filter for removing power line interference (50 Hz), the presented method can effectively remove power line interference and its harmonic wave components, and it completely reserves useful information contained in the sEMG signal. The linear change process of grip force at five grip speeds are studied. A simple and efficient method for muscle contraction state detection is used. Six supervised classification algorithms are used to identify these grip patterns and achieve a highest recognition rate of over 99%. Five supervised regression algorithms are applied to predict the value of grip force and establish the mapping relationship between sEMG and force. The experimental results demonstate that the method can simulate the linear change process of grip force well, and it achieves a minimum standard deviation of 0.14272.

Evaluation Land Use Cover Changes Over 29 Years in Papua Province of Indonesia Using Remote Sensing Data

Land use/cover change (LUCC) observation and determination have been extensively discussed in natural resources management, biodiversity and ecosystem preservation, land management also climate changes studies. An evaluation of the LUCC in Merauke, the easternmost city of Indonesia, was conducted to gain relevant information in agriculture and forestry based on historical data from remotely sensed land cover data. To obtain the historical dynamics of the LUCC, a supervised classification algorithm was implemented to the Landsat images of 1990, 1996, 2009, 2015, and 2019 as well as statistical analysis Wilcoxon signed-rank. This study aims to investigate and analyze the LUCC over a period of 29 years. The results show a significant decline in the forest area by around 7% in 2019. As compared to the LUCC in 1990, agriculture sectors such as dry-land agriculture, shrub-mixed dry land and paddy field gained an area higher than 50%, while the area of the water body is a declined by 0.61 % over the period. The evaluation of LUCC will help government development plans in the protection of commodities in this Province and the promotion for biomass and bioenergy resources.

Transfer learning for autonomous chatter detection in machining

Large-amplitude chatter vibrations are one of the most important phenomena in machining processes. It is often detrimental in cutting operations causing a poor surface finish and decreased tool life. Therefore, chatter detection using machine learning has been an active research area over the last decade. Three challenges can be identified in applying machine learning for chatter detection at large in industry: an insufficient understanding of the universality of chatter features across different processes, the need for automating feature extraction, and the existence of limited data for each specific workpiece-machine tool combination, e.g., when machining one-off products. These three challenges can be grouped under the umbrella of transfer learning, which is concerned with studying how knowledge gained from one setting can be leveraged to obtain information in new settings. This paper studies automating chatter detection by evaluating transfer learning of prominent as well as novel chatter detection methods. We investigate chatter classification accuracy using a variety of features extracted from turning and milling experiments with different cutting configurations. The studied methods include Fast Fourier Transform (FFT), Power Spectral Density (PSD), the Auto-correlation Function (ACF), and decomposition based tools such as Wavelet Packet Transform (WPT) and Ensemble Empirical Mode Decomposition (EEMD). We also examine more recent approaches based on Topological Data Analysis (TDA) and similarity measures of time series based on Discrete Time Warping (DTW). We evaluate transfer learning potential of each approach by training and testing both within and across the turning and milling data sets. Four supervised classification algorithms are explored: support vector machine (SVM), logistic regression, random forest classification, and gradient boosting. In addition to accuracy, we also comment on the automation potential of feature extraction for each approach which is integral to creating autonomous manufacturing centers. Our results show that carefully chosen time-frequency features can lead to high classification accuracies albeit at the cost of requiring manual pre-processing and the tagging of an expert user. On the other hand, we found that the TDA and DTW approaches can provide accuracies and F1-scores on par with the time-frequency methods without the need for manual preprocessing via completely automatic pipelines. Further, we discovered that the DTW approach outperforms all other methods when trained using the milling data and tested on the turning data. Therefore, TDA and DTW approaches may be preferred over the time-frequency-based approaches for fully automated chatter detection schemes. DTW and TDA also can be more advantageous when pooling data from either limited workpiece-machine tool combinations, or from small data sets of one-off processes.

COVIDX: Computer‐aided diagnosis of COVID‐19 and its severity prediction with raw digital chest X‐ray scans

BackgroundCoronavirus disease (COVID‐19) is a contagious infection caused by severe acute respiratory syndrome coronavirus‐2 (SARS‐COV‐2) and it has infected and killed millions of people across the globe.ObjectiveIn the absence or inadequate provision of therapeutic treatments of COVID‐19 and the limited convenience of diagnostic techniques, there is a necessity for some alternate spontaneous screening systems that can easily be used by the physicians to rapidly recognize and isolate the infected patients to circumvent onward surge. A chest X‐ray (CXR) image can effortlessly be used as a substitute modality to diagnose the COVID‐19.MethodIn this study, we present an automatic COVID‐19 diagnostic and severity prediction system (COVIDX) that uses deep feature maps of CXR images along with classical machine learning algorithms to identify COVID‐19 and forecast its severity. The proposed system uses a three‐phase classification approach (healthy vs unhealthy, COVID‐19 vs pneumonia, and COVID‐19 severity) using different conventional supervised classification algorithms.ResultsWe evaluated COVIDX through 10‐fold cross‐validation, by using an external validation dataset, and also in a real setting by involving an experienced radiologist. In all the adopted evaluation settings, COVIDX showed strong generalization power and outperforms all the prevailing state‐of‐the‐art methods designed for this purpose.ConclusionsOur proposed method (COVIDX), with vivid performance in COVID‐19 diagnosis and its severity prediction, can be used as an aiding tool for clinical physicians and radiologists in the diagnosis and follow‐up studies of COVID‐19 infected patients.AvailabilityWe made COVIDX easily accessible through a cloud‐based webserver and python code available at the site of google and the website of Github.

MALDI Mass Spectrometry Imaging for the Distinction of Adenocarcinomas of the Pancreas and Biliary Tree.

Pancreatic ductal adenocarcinoma and cholangiocarcinoma constitute two aggressive tumor types that originate from the epithelial lining of the excretory ducts of the pancreatobiliary tract. Given their close histomorphological resemblance, a correct diagnosis can be challenging and almost impossible without clinical information. In this study, we investigated whether mass spectrometric peptide features could be employed to distinguish pancreatic ductal adenocarcinoma from cholangiocarcinoma. Three tissue microarrays of formalin-fixed and paraffin-embedded material (FFPE) comprising 41 cases of pancreatic ductal adenocarcinoma and 41 cases of cholangiocarcinoma were analyzed by matrix-assisted laser desorption/ionization mass spectrometry imaging (MALDI-MSI). The derived peptide features and respective intensities were used to build different supervised classification algorithms: gradient boosting (GB), support vector machine (SVM), and k-nearest neighbors (KNN). On a pixel-by-pixel level, a classification accuracy of up to 95% could be achieved. The tentative identification of discriminative tryptic peptide signatures revealed proteins that are involved in the epigenetic regulation of the genome and tumor microenvironment. Despite their histomorphological similarities, mass spectrometry imaging represents an efficient and reliable approach for the distinction of PDAC from CC, offering a promising complementary or alternative approach to the existing tools used in diagnostics such as immunohistochemistry.

A Decision Tree Model for Analysis and Judgment of Lower Limb Movement Comfort Level.

To address the problem of ambiguity and one-sidedness in the evaluation of comprehensive comfort perceptions during lower limb exercise, this paper deconstructs the comfort perception into two dimensions: psychological comfort and physiological comfort. Firstly, we designed a fixed-length weightless lower limb squat exercise test to collect original psychological comfort data and physiological comfort data. The principal component analysis and physiological comfort index algorithm were used to extract the comfort index from the original data. Secondly, comfort degrees for each sample were obtained by performing K-means++ to cluster normalized comfort index. Finally, we established a decision tree model for lower limb comfort level analysis and determination. The results showed that the classification accuracy of the model reached 95.8%, among which the classification accuracy of the four comfort levels reached 95.2%, 97.3%, 92.9%, and 97.8%, respectively. In order to verify the advantages of this paper, the classification results of this paper were compared with the classification results of four supervised classification algorithms: Gaussian Parsimonious Bayes, linear SVM, cosine KNN and traditional CLS decision tree.

Underwater Image Detection for Cleaning Purposes; Techniques Used for Detection Based on Machine Learning

Abstract Serious problems are on the rise, especially in these current times. The world is facing too many environmental threats. Water pollution is one of the main issues threatening the future. In some parts of the world, the water’s surface is covered by mucilage, which is dangerous for both aquatic animals and humans. This article firstly defines mucilage and highlights the reasons for its production. Afterwards to tackle water pollution, cleaning systems using image detection with the help of machine learning supervised classification algorithms are highlighted. This paper showcases the machine learning and classification used as well as the best solution for convolutional neural network and region-based convolutional neural network methods.

Traffic and road conditions monitoring system using extracted information from Twitter

Congested roads and daily traffic jams cause traffic disturbances. A traffic monitoring system using closed-circuit television (CCTV) has been implemented, but the information gathered is still limited for public use. This research focuses on utilizing Twitter data to monitor traffic and road conditions. Traffic-related information is extracted from social media using text mining approach. The methods include Tweet classification for filtering relevant data, location information extraction, and geocoding in order to convert text-based location into coordinate information that can be deployed into Geographic Information System. We test several supervised classification algorithms in this study, i.e., Naïve Bayes, Random Forest, Logistic Regression, and Support Vector Machine. We experiment with Bag Of Words (BOW) and Term Frequency - Inverse Document Frequency (TF-IDF) as the feature representation. The location information is extracted using Named Entity Recognition (NER) and Part-Of-Speech (POS) Tagger. The geocoding is implemented using the ArcPy library. The best model for Tweet relevance classification is the Logistic Regression classifier with the feature combination of unigram and char n-gram, achieving an F1-score of 93%. The NER-based location extractor obtains an F1-score of 54% with a precision of 96%. The geocoding success rate for extracting the location information is 68%. In addition, a web-based visualization is also implemented in order to display traffic information using the spatial interface.

Towards automated in vivo parcellation of the human cerebral cortex using supervised classification of magnetic resonance fingerprinting residuals.

Accurate parcellation of the cerebral cortex in an individual is a guide to its underlying organization. The most promising in vivo quantitative magnetic resonance (MR)-based microstructural cortical mapping methods are yet to achieve a level of parcellation accuracy comparable to quantitative histology. We scanned 6 participants using a 3D echo-planar imaging MR fingerprinting (EPI-MRF) sequence on a 7T Siemens scanner. After projecting MRF signals to the individual-specific inflated model of the cortical surface, normalized autocorrelations of MRF residuals of vertices of 8 microstructurally distinct areas (BA1, BA2, BA4a, BA6, BA44, BA45, BA17, and BA18) from 3 cortical regions were used as feature vector inputs into linear support vector machine (SVM), radial basis function SVM (RBF-SVM), random forest, and k-nearest neighbors supervised classification algorithms. The algorithms' prediction performance was compared using: (i) features from each vertex or (ii) features from neighboring vertices. The neighborhood-based RBF-SVM classifier achieved the highest prediction score of 0.85 for classification of MRF residuals in the central region from a held-out participant. We developed an automated method of cortical parcellation using a combination of MR fingerprinting residual analysis and machine learning classification. Our findings provide the basis for employing unsupervised learning algorithms for whole-cortex structural parcellation in individuals.

An integrated framework to identify and map gullies in a Mediterranean region of Turkey

This research introduces a scientific methodology to identify areas affected by gully erosion using Geographic Object Based Image Analysis (GEOBIA) and Random Forest (RF) supervised machine learning. The GEOBIA and RF were applied in Besni district, which has a Mediterranean climate, of Adiyaman province in Turkey by including many factors in the model. Estimation Scale Parameter (ESPII) algorithm was used in the segmentation phase. The novelty of this study is the implementation of RF supervised classification algorithm to classify a large number of objects determined after the segmentation process, due to the large size of the study area. Therefore, open access data has been evaluated with high classification accuracy without the need for labor. Precision, Recall and F1-Score values were calculated using true positive (TP), true negative (TN), false positive (FP) and false negative (FN) values based on field observations and Google Earth images of the study area. The TP, TN, FP and FN values were 0.90, 0.95 and 0.92, respectively. In addition, a Kappa-index was calculated as 0.88. The gully erosion map obtained using aforementioned methodology can be used to take necessary measures to prevent further degradation and plan sustainable land uses.

Machine Learning: A Review on Supervised Classification Algorithms and their Applications to Optical Character Recognition in Indic Scripts

Machine learning is a common application of artificial intelligence, which gives machines the ability to learn from data automatically and improve with experience without being explicitly programmed. Supervised learning is one of two broad areas of machine learning that deal with the task of learning a function for a current model based on past data training on input-output pairs of examples. This function enables the model to predict future outcomes for new inputs. Regression and classification are two supervised machine learning problems. Classification is the most common task that intelligent systems perform most often. This review describes the classification algorithm’s working and its applications to optical character recognition in Indic script. Devanagari and Gurumukhi scripts are chosen for this attempt. The detailed study presented in this article will be helpful to researchers working in the field of optical character recognition to understand where to use which machine learning algorithm for best results.

Remote Sensing and Spatial Analysis for Land-Take Assessment in Basilicata Region (Southern Italy)

Land use is one of the drivers of land-cover change (LCC) and represents the conversion of natural to artificial land cover. This work aims to describe the land-take-monitoring activities and analyze the development trend in test areas of the Basilicata region. Remote sensing is the primary technique for extracting land-use/land-cover (LULC) data. In this study, a new methodology of classification of Landsat data (TM–OLI) is proposed to detect land-cover information automatically and identify land take to perform a multi-temporal analysis. Moreover, within the defined model, it is crucial to use the territorial information layers of geotopographic database (GTDB) for the detailed definition of the land take. All stages of the classification process were developed using the supervised classification algorithm support vector machine (SVM) change-detection analysis, thus integrating the geographic information system (GIS) remote sensing data and adopting free and open-source software and data. The application of the proposed method allowed us to quickly extract detailed land-take maps with an overall accuracy greater than 90%, reducing the cost and processing time.

Comparing Fire Extent and Severity Mapping between Sentinel 2 and Landsat 8 Satellite Sensors

Mapping of fire extent and severity across broad landscapes and timeframes using remote sensing approaches is valuable to inform ecological research, biodiversity conservation and fire management. Compiling imagery from various satellite sensors can assist in long-term fire history mapping; however, inherent sensor differences need to be considered. The New South Wales Fire Extent and Severity Mapping (FESM) program uses imagery from Sentinel and Landsat satellites, along with supervised classification algorithms, to produce state-wide fire maps over recent decades. In this study, we compared FESM outputs from Sentinel 2 and Landsat 8 sensors, which have different spatial and spectral resolutions. We undertook independent accuracy assessments of both Sentinel 2 and Landsat 8 sensor algorithms using high-resolution aerial imagery from eight training fires. We also compared the FESM outputs from both sensors across 27 case study fires. We compared the mapped areas of fire severity classes between outputs and assessed the classification agreement at random sampling points. Our independent accuracy assessment demonstrated very similar levels of accuracy for both sensor algorithms. We also found that there was substantial agreement between the outputs from the two sensors. Agreement on the extent of burnt versus unburnt areas was very high, and the severity classification of burnt areas was typically either in agreement between the sensors or in disagreement by only one severity class (e.g., low and moderate severity or high and extreme severity). Differences between outputs are likely partly due to differences in sensor resolution (10 m and 30 m pixel sizes for Sentinel 2 and Landsat 8, respectively) and may be influenced by landscape complexity, such as terrain roughness and foliage cover. Overall, this study supports the combined use of both sensors in remote sensing applications for fire extent and severity mapping.

A machine learning classification methodology for Raman Hyperspectral imagery based on auto‐encoders

AbstractRaman spectroscopy is a very sensitive, non‐destructive optical technique that is able to characterize the composition of a material. Although it is a very useful technique, its signals are rather complex and therefore require a lot of knowledge and skills to understand their meaning. This complexity increases further, when it is combined with spatial mapping, creating a three‐dimensional dataset named Raman hyperspectral imagery. The common methodology of analysing such a dataset is to use a generic supervised classification algorithm (i.e., non‐negative least square) combined with reference data. If reference data are not available, unsupervised algorithms are used to perform spectral clusterization of similar areas. This is typically achieved using standard PCA and k‐means methods. The most common problem of processing that significantly affects the results of analysis is that a raw Raman Signal consists of three sources: Raman scattering, photoluminescence and noise. The contribution weights of those components are not standard and are changing based on the sample's chemistry, which makes it very difficult to be solved. This creates the necessity for pre‐processing methods that can split this raw signal to its components. The goals of this project are to improve the unmixing performance using AI and to achieve a successful split of the raw signal into its three components. Through this study, various unsupervised methods are explored to identify the method that fits best to our microscopic RHSI data sets. Our focus is on using autoencoders for the data analysis and on a Gaussian smoothing filter combined with polynomial fittings to split the data. Results show that the selection of a classification methodology depends greatly on the application, but when a large RHSI dataset is available, the newly proposed splitting technique combined with an autoencoder classification presents the best solution.

Fourier transform infrared spectroscopy and chemometrics for the discrimination of animal fur types

Rapid and reliable animal fur identification has remained a challenge for customs inspection. The accurate distinction between fur types has a significant meaning in implementing the correct tariff policy. A variety of analytical methods have been applied to work on distinguishing animal fur types, with tools of microscopy, molecular testing, mass spectrometry, Fourier transform infrared spectroscopy (FTIR), and Raman spectroscopy. In this research, the capability of attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR) combined with pattern recognition methods was investigated for the discrimination of animal fur in six types. This work was to explore the non-destructive application of ATR-FTIR technique in discriminant analysis of animal fur. All spectra were collected by ATR-FTIR of the wavenumber ranging from 4000 to 650 cm−1. Data pretreatments included moving average smoothing and multiplicative scatter correction (MSC). Four supervised classification algorithms were chosen to categorize the types of fur: soft independent modeling of class analogy (SIMCA), principal component analysis linear discriminant analysis (PCA-LDA), partial least squares discriminant analysis (PLS-DA), least squares support vector machine (LS-SVM). PLS-DA and LS-SVM were both effective approaches, with a 100% classification accuracy rate. The accuracy of PCA-LDA and SIMCA was 98.33% and 99.44%, respectively. Furthermore, LS-SVM model obtained using Monte-Carlo sampling method also obtained 100% prediction accuracy, while all other methods produced misclassification. LS-SVM corrected the non-linearities for the animal fur FTIR data but also remarkably improved the prediction performance level. The results of this study revealed that the combination of ATR-FTIR and chemometrics has a huge potential for animal fur discrimination.

Machine Learning for Outcome Prediction in First-Line Surgery of Prolactinomas.

BackgroundFirst-line surgery for prolactinomas has gained increasing acceptance, but the indication still remains controversial. Thus, accurate prediction of unfavorable outcomes after upfront surgery in prolactinoma patients is critical for the triage of therapy and for interdisciplinary decision-making.ObjectiveTo evaluate whether contemporary machine learning (ML) methods can facilitate this crucial prediction task in a large cohort of prolactinoma patients with first-line surgery, we investigated the performance of various classes of supervised classification algorithms. The primary endpoint was ML-applied risk prediction of long-term dopamine agonist (DA) dependency. The secondary outcome was the prediction of the early and long-term control of hyperprolactinemia.MethodsBy jointly examining two independent performance metrics – the area under the receiver operating characteristic (AUROC) and the Matthews correlation coefficient (MCC) – in combination with a stacked super learner, we present a novel perspective on how to assess and compare the discrimination capacity of a set of binary classifiers.ResultsWe demonstrate that for upfront surgery in prolactinoma patients there are not a one-algorithm-fits-all solution in outcome prediction: different algorithms perform best for different time points and different outcomes parameters. In addition, ML classifiers outperform logistic regression in both performance metrics in our cohort when predicting the primary outcome at long-term follow-up and secondary outcome at early follow-up, thus provide an added benefit in risk prediction modeling. In such a setting, the stacking framework of combining the predictions of individual base learners in a so-called super learner offers great potential: the super learner exhibits very good prediction skill for the primary outcome (AUROC: mean 0.9, 95% CI: 0.92 – 1.00; MCC: 0.85, 95% CI: 0.60 – 1.00). In contrast, predicting control of hyperprolactinemia is challenging, in particular in terms of early follow-up (AUROC: 0.69, 95% CI: 0.50 – 0.83) vs. long-term follow-up (AUROC: 0.80, 95% CI: 0.58 – 0.97). It is of clinical importance that baseline prolactin levels are by far the most important outcome predictor at early follow-up, whereas remissions at 30 days dominate the ML prediction skill for DA-dependency over the long-term.ConclusionsThis study highlights the performance benefits of combining a diverse set of classification algorithms to predict the outcome of first-line surgery in prolactinoma patients. We demonstrate the added benefit of considering two performance metrics jointly to assess the discrimination capacity of a diverse set of classifiers.