Foundation of Machine Learning-Based Data Classification Techniques for Health Care

An intrusion detection system (IDS) is an important protection instrument for detecting complex network attacks. Various machine learning (ML) or deep learning (DL) algorithms have been proposed for implementing anomaly-based IDS (AIDS). Our review of the AIDS literature identifies some issues in related work, including the randomness of the selected algorithms, parameters, and testing criteria, the application of old datasets, or shallow analyses and validation of the results. This paper comprehensively reviews previous studies on AIDS by using a set of criteria with different datasets and types of attacks to set benchmarking outcomes that can reveal the suitable AIDS algorithms, parameters, and testing criteria. Specifically, this paper applies 10 popular supervised and unsupervised ML algorithms for identifying effective and efficient ML-AIDS of networks and computers. These supervised ML algorithms include the artificial neural network (ANN), decision tree (DT), k-nearest neighbor (k-NN), naive Bayes (NB), random forest (RF), support vector machine (SVM), and convolutional neural network (CNN) algorithms, whereas the unsupervised ML algorithms include the expectation-maximization (EM), k-means, and self-organizing maps (SOM) algorithms. Several models of these algorithms are introduced, and the turning and training parameters of each algorithm are examined to achieve an optimal classifier evaluation. Unlike previous studies, this study evaluates the performance of AIDS by measuring the true positive and negative rates, accuracy, precision, recall, and F-Score of 31 ML-AIDS models. The training and testing time for ML-AIDS models are also considered in measuring their performance efficiency given that time complexity is an important factor in AIDSs. The ML-AIDS models are tested by using a recent and highly unbalanced multiclass CICIDS2017 dataset that involves real-world network attacks. In general, the k-NN-AIDS, DT-AIDS, and NB-AIDS models obtain the best results and show a greater capability in detecting web attacks compared with other models that demonstrate irregular and inferior results.

In this paper, a novel unsupervised machine learning (ML) algorithm is presented for the expeditious RF fingerprinting of LoRa modulated chirps. Identification based on received signal strength indicator (RSSI) alone is unlikely to yield a robust means for sensor authentication within critical infrastructure deployment. Here, an unsupervised ML algorithm is used to rapidly train an artificial neural network (ANN) matrix creating self-organizing maps (SOMs) for each authentic transmitter and a potential rogue node. A general classifier can be trained on the SOMs for precisely profiling each transmitter as either genuine or rogue. By means of experimental validation, this methodology demonstrated cent-percent success in recognizing each transmitter, either being a real or a rogue node.

Directed Graph based models of a blockchain that capture accounts as nodes and transactions as edges, evolve over time. This temporal nature of a blockchain model enables us to understand the behavior (malicious or benign) of the accounts. Predictive classification of accounts as malicious or benign could help users of the permissionless blockchain platforms to operate in a secure manner. Motivated by this, we introduce temporal features such as burst and attractiveness on top of several already used graph properties such as the node degree and clustering coefficient. Using identified features, we train various Machine Learning (ML) models and identify the algorithm that performs the best in detecting malicious accounts. We then study the behavior of the accounts over different temporal granularities of the dataset before assigning them malicious tags. For the Ethereum blockchain, we identify that for the entire dataset—the ExtraTreesClassifier performs the best among supervised ML algorithms. On the other hand, using cosine similarity on top of the results provided by unsupervised ML algorithms such as K-Means on the entire dataset, we were able to detect 554 more suspicious accounts. Further, using behavior change analysis for accounts, we identify 814 unique suspicious accounts across different temporal granularities.

This paper analyses the recent evolution of the electricity price of one of the EastEuropean countries’ Balancing Markets (BM) – Romania, aiming to understand the prices trend and predict them in the current economic and geopolitical context. This is especially important as the electricity producers have to allocate their output between wholesale electricity market, ancillary services markets and BM targeting to maximize value and achieve a sustainable economic development. Therefore, in this paper, we propose an AI-powered electricity price forecast using several types of standout Machine Learning (ML) algorithms such as classifiers and regressors to predict the electricity price on BM. This approach, consisting of two steps, identifies the imbalance sign and significantly enhances the performance of the price forecast. The proposed method offers valuable insights into the market participants’ trading opportunities using two prediction solutions. The first prediction solution consists of averaging the results of five ensemble ML algorithms. The second one consists in weighting the results of the five forecasting ML algorithms using either a linear regression or a decision tree algorithm. Thus, we propose to combine supervised and unsupervised ML algorithms and find the fundamentals for creating optimal bidding strategies for electricity market players.

In the actual panorama of machine learning (ML) algorithms, the issue of the real-time information extraction/classification/manipulation/analysis of data streams (DS) is acquiring an ever-growing relevance. They arrive generally at high speed and always require an unsupervised real-time analysis for individuating long-range and higher order correlations among data that are continuously changing over time (phase transitions). This emphasizes the infinitary character of the issue, i.e., the continuous change of the signifying number of degrees of freedom characterizing the statistical representation function, challenging the classical ML algorithms, both in their classical and quantum versions, as far as all are based on the (stochastic) search for the global minimum of some cost/energy function. The physical analogue must be studied in the realm of quantum field theory (QFT) for dissipative systems as biological and neural systems, which are able to map between different phases of quantum fields, using the formalism of the Bogoliubov transform (BT). By applying the BT in a reversed way, on the system-thermal bath energetically balanced states, it is possible to define the powerful computational tool of the “doubling of the degrees of freedom” (DDF), making the choice of the signifying finite number of the degrees of freedom dynamic and then automatic, so to suggest a different class of unsupervised ML algorithms for solving the DS issue.

Chronic urticaria (CU) comes as chronic spontaneous urticaria (CSU) and chronic inducible urticaria (CIndU).1 Across its types and subtypes, CU is a heterogeneous disease that has different phenotypes with distinct clinical characteristics and different endotypes with distinct underlying pathophysiological mechanisms.2, 3 It may be possible that subtypes of CU patients exhibit distinct phenotypic disease signatures that can point to differences in what drives their condition and in their response to treatments. Cluster analysis is a popular unsupervised machine learning (ML) method for discovering previously undetected data patterns.4 ML-based cluster analysis has been used in several diseases for the identification and characterization of patient subgroups.5, 6 As of now, no study has attempted to identify CU subtypes with this method. Here, we performed a proof-of-concept study to test whether cluster analysis using ML algorithms can identify subgroups of CU patients based on clinical and routine laboratory characteristics. We retrospectively analyzed the medical charts of a cohort of 431 CU patients. Institutional review board was obtained, and due to retrospective nature of the study, patient consent was not required. ML-based k-means clustering with principal component silhouette analyses (PCA) and use of the elbow method of dimensionally reduced data showed 4 clusters of CU patients, with a homogeneous balance between the clusters and the selected evaluation metrics (methods are provided in supplementary material) (Figure S3). Clustering analyses with PCA resulted in more meaningful clusters than without and supported the positive impact of reduced dimensions, and cluster number identified. Cluster characteristics and comparisons identified clinically distinct patient subgroups (Table 1). Cluster 1 (The “CIndU only” cluster) was the smallest cluster and consisted of all and only CIndU patients who did not have comorbid CSU. Of all clusters, cluster 1 patients had the highest age [median 42 (28–51) years], the shortest duration of disease [12 (5–96) months], and the lowest IgE levels [74.6 (35.1–188.5) IU/ml]. Cluster 2 (The “high IgE” cluster) was the largest cluster. All patients had CSU, and half of them had comorbid CIndU. Cluster 2 patients, on average, had the highest IgE levels [132 (56.4–271.5) IU/ml], the highest rate of comorbid atopic dermatitis (7.7%), and the lowest rate of ANA and IgG-anti-TPO positivity (1.4% and 4.2%, respectively). Cluster 3 (The “autoimmune” cluster) had the highest percentage of women (92%) in all clusters. All patients had CSU, and more than half also had CIndU (56.3%). Three of four patients (77.3%) had angioedema, the highest percentage of any cluster. Cluster 3 patients also had the second-lowest IgE levels (84.2 IU/ml) of any CSU cluster and the highest rates of IgG-anti-TPO and ANA positivity across all clusters (39.8% and 52.3%, respectively). Cluster 4 (The “high comorbidity” cluster) consisted only of CSU patients, and comorbid CIndU was rare (14.3%). The defining characteristics of patients in this cluster, the high comorbidity cluster, were their high rates of hypertension (74%), diabetes mellitus (62%), and hypothyroidism (38%), each at least twice as high as in any other cluster. The results of our study provide proof of concept that the use of unsupervised ML algorithms can identify meaningful and distinct groups of patients with CU and cluster CU into four different and distinct subtypes. Three of these four clusters are remarkably similar to how patients with CU are classified in real life, that is, as having CIndU or CSU as their primary form of CU and, in the latter, as having autoimmune or autoallergic CSU (Figure 1). This suggests that ML-based algorithms can be used to establish patient signatures, which may then be used to better characterize relevant and distinct pathomechanisms of CU subgroups. This, in turn, will allow us to better manage CU, by optimizing the use of available treatments and guiding the development of new and better ones. This project benefitted from the support (non-financial) of the GA2LEN network of urticaria centers of reference and excellence (UCARE, www.ga2len-ucare.com). MT has no relevant conflict of interest in relation to this work. Outside of it, MT is or recently was a speaker and/or advisor for Novartis. RE has no relevant conflict of interest in relation to this work. Outside of it, RE is or recently was a speaker and/or advisor for Novartis. EZ, YT, MA, and AG have no conflict of interest. MM has no relevant conflicts of interest in relation to this work. Outside of it, MM is or recently was a speaker and/or advisor for and/or has received research funding from Allakos, Amgen, Aralez, ArgenX, AstraZeneca, Celldex, Centogene, CSL Behring, FAES, Genentech, GIInnovation, Innate Pharma, Kyowa Kirin, Leo Pharma, Lilly, Menarini, Moxie, Novartis, Roche, Sanofi/Regeneron, Third HarmonicBio, UCB, and Uriach. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

ObjectivesThe five-repetition sit-to-stand (5R-STS) test was designed to capture objective functional impairment (OFI), and thus provides an adjunctive dimension in patient assessment. It is conceivable that there are different subsets of patients with OFI and degenerative lumbar disease. We aim to identify clusters of objectively functionally impaired individuals based on 5R-STS and unsupervised machine learning (ML).MethodsData from two prospective cohort studies on patients with surgery for degenerative lumbar disease and 5R-STS times of ≥ 10.5 s—indicating presence of OFI. K-means clustering—an unsupervised ML algorithm—was applied to identify clusters of OFI. Cluster hallmarks were then identified using descriptive and inferential statistical analyses.ResultsWe included 173 patients (mean age [standard deviation]: 46.7 [12.7] years, 45% male) and identified three types of OFI. OFI Type 1 (57 pts., 32.9%), Type 2 (81 pts., 46.8%), and Type 3 (35 pts., 20.2%) exhibited mean 5R-STS test times of 14.0 (3.2), 14.5 (3.3), and 27.1 (4.4) seconds, respectively. The grades of OFI according to the validated baseline severity stratification of the 5R-STS increased significantly with each OFI type, as did extreme anxiety and depression symptoms, issues with mobility and daily activities. Types 1 and 2 are characterized by mild to moderate OFI—with female gender, lower body mass index, and less smokers as Type I hallmarks.ConclusionsUnsupervised learning techniques identified three distinct clusters of patients with OFI that may represent a more holistic clinical classification of patients with OFI than test-time stratifications alone, by accounting for individual patient characteristics.

Surgical site infection (SSI) following a neurosurgical operation is a complication that impacts morbidity, mortality, and economics. Currently, machine learning (ML) algorithms are used for outcome prediction in various neurosurgical aspects. The implementation of ML algorithms to learn from medical data may help in obtaining prognostic information on diseases, especially SSIs. The purpose of this study was to compare the performance of various ML models for predicting surgical infection after neurosurgical operations. A retrospective cohort study was conducted on patients who had undergone neurosurgical operations at tertiary care hospitals between 2010 and 2017. Supervised ML algorithms, which included decision tree, naive Bayes with Laplace correction, k-nearest neighbors, and artificial neural networks, were trained and tested as binary classifiers (infection or no infection). To evaluate the ML models from the testing data set, their sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), as well as their accuracy, receiver operating characteristic curve, and area under the receiver operating characteristic curve (AUC) were analyzed. Data were available for 1471 patients in the study period. The SSI rate was 4.6%, and the type of SSI was superficial, deep, and organ/space in 1.2%, 0.8%, and 2.6% of cases, respectively. Using the backward stepwise method, the authors determined that the significant predictors of SSI in the multivariable Cox regression analysis were postoperative CSF leakage/subgaleal collection (HR 4.24, p < 0.001) and postoperative fever (HR 1.67, p = 0.04). Compared with other ML algorithms, the naive Bayes had the highest performance with sensitivity at 63%, specificity at 87%, PPV at 29%, NPV at 96%, and AUC at 76%. The naive Bayes algorithm is highlighted as an accurate ML method for predicting SSI after neurosurgical operations because of its reasonable accuracy. Thus, it can be used to effectively predict SSI in individual neurosurgical patients. Therefore, close monitoring and allocation of treatment strategies can be informed by ML predictions in general practice.

To detect unusual infusion alerting patterns using machine learning (ML) algorithms as a first step to advance safer inpatient intravenous administration of high-alert medications. We used one year of detailed propofol infusion data from a hospital. Interpretable and clinically relevant variables were feature engineered, and data points were aggregated per calendar day. A univariate (maximum times-limit) moving range (mr) control chart was used to simulate clinicians' common approach to identifying unusual infusion alerting patterns. Three different unsupervised multivariate ML-based anomaly detection algorithms (Local Outlier Factor, Isolation Forest, and k-Nearest Neighbors) were used for the same purpose. Results from the control chart and ML algorithms were compared. The propofol data had 3,300 infusion alerts, 92% of which were generated during the day shift and seven of which had a times-limit greater than 10. The mr-chart identified 15 alert pattern anomalies. Different thresholds were set to include the top 15 anomalies from each ML algorithm. A total of 31 unique ML anomalies were grouped and ranked by agreeability. All algorithms agreed on 10% of the anomalies, and at least two algorithms agreed on 36%. Each algorithm detected one specific anomaly that the mr-chart did not detect. The anomaly represented a day with 71 propofol alerts (half of which were overridden) generated at an average rate of 1.06 per infusion, whereas the moving alert rate for the week was 0.35 per infusion. These findings show that ML-based algorithms are more robust than control charts in detecting unusual alerting patterns. However, we recommend using a combination of algorithms, as multiple algorithms serve a benchmarking function and allow researchers to focus on data points with the highest algorithm agreeability. Unsupervised ML algorithms can assist clinicians in identifying unusual alert patterns as a first step toward achieving safer infusion practices.

Performance evaluation of machine learning techniques in predicting cumulative absolute velocity

Machine learning analysis of bleeding status in venous thromboembolism patients

Recently, various sophisticated methods, including machine learning and artificial intelligence, have been employed to examine health-related data. Medical professionals are acquiring enhanced diagnostic and treatment abilities by utilizing machine learning applications in the healthcare domain. Medical data have been used by many researchers to detect diseases and identify patterns. In the current literature, there are very few studies that address machine learning algorithms to improve healthcare data accuracy and efficiency. We examined the effectiveness of machine learning algorithms in improving time series healthcare metrics for heart rate data transmission (accuracy and efficiency). In this paper, we reviewed several machine learning algorithms in healthcare applications. After a comprehensive overview and investigation of supervised and unsupervised machine learning algorithms, we also demonstrated time series tasks based on past values (along with reviewing their feasibility for both small and large datasets).

P125. Development of a novel ensemble machine learning algorithm for prediction of complications and readmission after anterior cervical spinal fusion

Background Interpreting patient phenotypes is a challenge when screening for hypertrophic cardiomyopathy (HCM). Machine learning (ML) can potentially help with advanced data integration - combining information contained in whole-cardiac cycle echo deformation and velocity profiles with standard clinical variables. The aim is to apply an ML approach to integrate whole cardiac cycle echo data with clinical variables to explore HCM phenotypes. Methods The cohort consisted of 138 participants from two centres: HCM patients (n=91) and relatives (n=47). Echocardiography was performed, whereas magnetic resonance and genetic testing in 48% and 82%, respectively. Whole cardiac cycle echo data (mitral and aortic velocity profiles, and six regional left ventricular (LV) deformation curves) were combined with clinical variables (age, sex, heart rate, e' medial and e' lateral) and used as the ML input. An unsupervised ML algorithm created a representative space where participants were positioned based on integrated data, blinded to disease status. Clustering was used to determine phenogroups and estimate the average characteristics. Data on family history (FHx), genotype, arrhythmias or syncope, implantable cardioverter-defibrillators (ICD), and late gadolinium enhancement (LGE) were used to interpret the phenogroups. As the LA diameter was not available in the dataset, the HCM risk for sudden cardiac death (SCD) was not calculated, however, the Table shows relevant variables to infer clinical risk. Results Clustering divided the participants into 6 phenogroups (P1–6) (Figure). Average echo profiles are shown in the Figure, while the clinical data in the Table. P1/2 was defined by symptomatic patients with a high prevalence of positive genotypes, a positive FHx of SCD, and a burden of comorbidities. Echo findings showed pronounced structural/functional remodeling, and P1 was associated with severe septal hypertrophy and outflow tract obstruction. The high prevalence of ICD devices defined P1/2 as high risk groups. In comparison, patients in P3/4 were younger, with milder LV hypertrophy, but still considerable functional impairment. P3 had a higher burden of FHX and a higher prevalence of pathogenic mutations, whereas P4 a higher incidence of hypertension, high heart rate, mitral inflow fusion and findings of LGE. Finally, P5/6 consisted of younger individuals, predominantly HCM relatives, with a mild phenotype and, thus, low inferred risk. As expected, the majority of patients with the genetic variants of undetermined significance were located in P5. Conclusion ML can help derive clinically interpretable phenotypes in HCM based on the automated integration of whole cardiac cycle deformation and velocity data with conventional clinical parameters. The derived phenogroups correspond with established risk profiles in HCM. An expanded dataset is needed to enable further exploration of the phenotype-genotype relations and to define prognostic value. Funding Acknowledgement Type of funding sources: Public grant(s) – EU funding. Main funding source(s): This work was supported by the Horizon 2020 European Commission Project H2020-MSCA-ITN

The insufficient loading of vehicles, which leads to a low logistics load factor is a common problem in the logistics industry. This study addresses this issue by utilizing actual shipment data from an automotive company. An effective method has been proposed to improve the company’s logistics efficiency through a scenario-based approach. Two real- world scenarios were developed to enhance vehicle loading performance. Machine learning algorithms were employed to evaluate the shipment performance of these scenarios. For the study, a dataset was generated from the company’s ERP system and real-world shipment data. Since this is a real-world problem, the dataset consisted of unlabeled data. To solve this problem, both supervised and unsupervised learning algorithms were applied. First, unsupervised clustering algorithms were used to group the shipment performance based on similarities. Then, supervised learning algorithms were utilized to classify the data within each group. The average cost was used to evaluate the clusters obtained through the unsupervised methods, while classification performance was measured using supervised machine learning techniques. The scenario-based approach has significantly improved the performance of the shipments as it shows the changes in load factor more clearly. In the actual case, only %25.7 of shipments were high performing, while this percentage gradually increased to %98.4 in the scenarios. The results show that optimizing the load factor makes the transports more efficient and balanced.