7 research outputs found
A Novel Spectral Feature Selection Method Based on Binary Genetic Algorithm for Efficient Detection of Endometrial and Ovarian Cancers: Preliminary Results
No full textDeep feature extraction, dimensionality reduction, and classification of medical images using combined deep learning architectures, autoencoder, and multiple machine learning models
No full textAccurate analysis and classification of medical images are essential factors in clinical decision-making and patient care. A novel comparative approach for medical image classification is proposed in this study. This new approach involves several steps: deep feature extraction, which extracts the informative features from medical images; concatenation, which concatenates the extracted deep features to form a robust feature vector; dimensionality reduction with autoencoder, which reduces the dimensionality of the feature vector by transforming it into a different feature space with a lower dimension; and finally, these features obtained from all these steps were fed into multiple machine learning classifiers (SVM, KNN, linear DA, and ANN) for the classification purpose. The study is performed to conduct a comparative analysis, aiming to evaluate the individual impact of each step within the proposed methodology and also assess the performance of each implemented classifier in order to find a best pipeline for medical image classification. The effectiveness of the proposed approach is assessed using two different medical image datasets. The performance assessment for the classifiers implemented is achieved using overall accuracy, sensitivity, and specificity metrics. The findings show that the linear DA classifier preceded by deep feature extraction, concatenation, and dimensionality reduction reveals itself to be a very efficient pipeline for accurate classification of medical images by utilizing a very small number of features
Dynamic Predictive Models for Side Effects Following Cancer or Cancer Treatment: A Systematic Review
No full textBackground: Advances in cancer treatments such as surgery, radiotherapy and chemotherapy have increased patient survival rates. However, these treatments often result in complications or late side-effects in patients. Accurate predictions of these side-effects are important to optimize post-treatment care focused on targeted interventions for prevention or management. Dynamic predictive models, which are specifically tailored to adapt to longitudinal patient data, offer an innovative approach to updating patient risk in light of new data. This systematic review aims to summarize the application of dynamic predictive models in predicting cancer treatment-related complications and to synthesize techniques and algorithms used in developing and validating these models.Methods: This review was conducted following the PRISMA guidelines. A systematic search was performed across multiple databases including Scopus, Web of Science, PubMed, and IEEE Xplore to identify studies that have employed dynamic predictive models for cancer or cancer treatment related complications or side-effects. The keywords used included “dynamic prediction”, “predictive models”, “treatment side effects”, “cancer treatment” and “over time”. Studies were included if they have employed longitudinal or time-varying data to update predictions over time, reflecting the incorporation of new data at multiple time points.Preliminary Results: A total of 506 studies were initially screened, resulting in the inclusion of 13 articles. Modelling techniques varied, including statistical models such as Coxproportional hazards and machine learning-based models like Long Short-Term Memory (LSTM) employed for handling time series data. The included studies were found to cover various cancer types, with prostate and head and neck cancers being the most common. Treatment types included surgery, radiation therapy, chemotherapy and hormone therapy. Predicted complications ranged from biochemical recurrence to patient-reported outcomes such as voice impairment. Each model employed different strategies for dynamically incorporating follow-up data. Most studies were from 2020–2024, reflecting a recent focus on dynamic models.Conclusion: Despite their versatility, dynamic models are not often used in oncology applications. This review highlights the diverse applications of dynamic models in predicting cancer or cancer treatment-related complications or side effects over time. These models showcase significant potential for improving post-treatment care by updating predictions as more data becomes available
One size does not fit all: a scoping review on study population diversity in studies assessing the validity of consumer wearables for measuring vital signs
No full textBackground: Consumer wearables provide promising opportunities for early detection and prevention of disease through continuous remote monitoring of vital signs. However, before implementing consumer wearables into research and healthcare settings their validity needs to be assessed. As user characteristics like sex, BMI and skin tone could influence the validity of wearable sensors, it is important that different categories of these characteristics are represented in the studies assessing wearable validity. So far, no study has investigated this representation, which makes it difficult to identify gaps in representation and provide recommendations for future studies.Objective: This scoping review aims to provide an overview of the representation and reporting of study population characteristics in studies that assess the validity of consumer wearables for measuring vital signs. These vital signs include heart rate, heart rate variability, blood pressure, blood oxygen saturation, respiratory rate and body temperature. Methods: A search was conducted in Scopus, PubMed and IEEE Xplore databases in June 2024. Publications were eligible if they assessed the validity of one or more consumer wearables for one or more vital signs in humans compared to a reference method. Publications were excluded if: 1) the wearable assessed was a chest strap or not suitable for wear during daily activities; 2) the study included only one participant. Results: After screening, a total of 161 studies were included. By extracting and synthesizing the study population characteristics for these studies, we will present gaps in reporting and representation of user characteristics including age, BMI and skin tone. Based on these findings, recommendations for reporting and representation for future studies will be provided. Conclusion: This scoping review will contribute to a more inclusive transition to home-based healthcare by identifying gaps in reporting and representation of study population characteristics in consumer wearable validation studies. <br/
Reporting, representation and subgroup analysis in studies assessing consumer wearable validity: a scoping review
Get PDFBackground: Consumer wearables provide promising opportunities for early detection and prevention of disease through continuous remote monitoring of vital signs. In many cases, however, the validity of the measurements from these devices is unknown and thus has to be assessed. User characteristics like sex, age, BMI and skin tone could influence the validity of wearable sensors. Therefore, it is important that studies assessing this validity: 1) report the distribution of these characteristics in their study population; 2) ensure representation of different categories of these characteristics within the study and across studies; 3) perform appropriate subgroup analysis to investigate differences in validity outcomes between categories. It is currently unclear to what extend these user characteristics are reported and represented in consumer wearable validation studies and which analysis methods are used to investigate the influence of these characteristics. This scoping review aims to map the current reporting, representation, subgroup analysis results and subgroup analysis methodology to provide recommendations for future validation studies.Methods: A literature search was conducted in Scopus, PubMed and IEEE Xplore in June 2024. Publications were eligible if they assessed the validity of consumer wearable vital sign measurements expressed as the agreement with a reference method. After duplicate removal, 551 publications were screened for eligibility of which 164 were included. The percentage of people in specific age categories, BMI categories and Fitzpatrick skin tone scale categories was estimated based on reported means and standard deviations. Additionally, the methods and results from 26 included publications investigating the influence of these user characteristics on validity were compiled.Findings: Of the 164, only 14% reported skin tone and 50% reported BMI. Almost all studies (95%) reported sex and age. When it comes to representation, the median percentage of older adults [0%], people with a BMI <18.5 kg/m2 [4%] or ≥30 kg/m2 [6%], and people with Fitzpatrick skin tone category 5 [1%] or 6 [0%] in a study was low, compared to the other groups. When it comes to the subgroup analysis, some studies reported a significant effect of sex (2/11), age (2/11), BMI (2/8) or Fitzpatrick scale (2/8) on heart rate measurement validity. However, among these studies contradicting results were found. Variation in sample size, the validity outcome used (e.g., absolute vs relative error) and the physical activity engaged in during the studies likely contributed to the inconsistency in results.Discussion: The results of this scoping review indicate gaps in the reporting and representation of user characteristics in studies. Additionally there are inconsistencies in the results of studies investigating the influence of these characteristics on validity. To close the current gaps in reporting and representation, reporting of BMI and skin tone and the representations of older adults, people with BMI <18.5 kg/m2 or ≥30 kg/m2 and people with darker skin tones should be increased. To reduce heterogeneity in results when it comes to the influence of user characteristics on validity, future validation studies should include multiple validity outcomes in the subgroup analysis, as well as investigate different physical activity levels separately
Unsupervised Detection of Postoperative Complications in Home-Monitored Patients: Preliminary Results
No full textWearable sensors enable remote, continuous patient monitoring at home, offering a promising approach for early detection of postoperative complications. However, analyzing continuous long-term physiological data remains challenging, particularly in the absence of precisely labeled deterioration events. Unsupervised change point detection methods can address this issue by identifying physiological deviations without requiring predefined event labels. This study investigates the feasibility of using a Long-Short-Term Memory (LSTM) autoencoder for detecting postoperative complications from continuous heart rate and respiration rate data using a wearable patch sensor while monitoring patients in their homes. The autoencoder was applied to identify physiological deviations that may indicate potential complications after major abdominal oncological surgeries in ten patients. The model was trained on data from seven patients to recognize deviations from normal physiological patterns and evaluated on three patients. The proposed model detected change points preceding the clinically documented complication time in two test patients, identifying these deteriorations an average of 3.25 hours earlier than the standard Remote Early Warning Score (REWS) alarm system. These findings suggest that LSTM autoencoder-based change point detection could be a valuable tool for identifying postoperative complications early in remote patient monitoring settings, to support timely intervention and potentially improving patient outcomes
Unsupervised Change Point Detection for Early Complication Identification in Post-Surgical Oncology Patients
Get PDFBackground: Postoperative complications following major abdominal surgeries are associated with high morbidity and prolonged hospital stays. Early detection of physiological deterioration is crucial for timely intervention. Traditional monitoring methods, which rely on periodic vital signs measurements and fixed thresholds, may fail to identify early signs of deterioration or can generate false alarms. Continuous physiological tracking presents a promising solution, especially with increasing trend of remote monitoring. However, it requires advanced analytics to differentiate between clinically relevant changes and normal fluctuations. . To address this challenge, we propose an unsupervised deep learning approach using an LSTM (Long Short-Term Memory) autoencoder to analyze continuous physiological data and detect potential complications early.Methods: Physiological data were collected from 30 patients who underwent oncological colorectal or pancreatic resections and were monitored with HealthDot sensor, a wearable patch sensor, for 14 days postoperatively at their homes. The sensor recorded various physiological parameters, including heart rate and respiratory rate. A subset of 10 patients (5 with complications, 5 without) was selected for model evaluation. An LSTM autoencoder was trained on continuous heart rate and respiratory rate data to learn physiological patterns and detect change points associated with potential complications. LSTMs are well-suited for time series data because they can maintain information over long intervals, making them ideal for scenarios where long-term trends that affect future values. The LSTM autoencoder aimed to reconstruct the input signal accurately based on normal recovery patterns. The reconstruction errors were used to identify change points, with a dynamic thresholding method applied for their detection. The Time-to-Detection (TTD) was calculated as the time difference between the first detected change point and the first clinically documented complication.Findings: The proposed unsupervised deep learning model identified physiological deteriorations up to 50.12 hours and 49.74 hours before the first clinically documented complications in two test patients. These results highlight the potential of combining LSTM-based autoencoder with dynamic thresholding for early detection of postoperative complications. However, change points were also observed in patients without recorded complications, which indicate the need for further refinement to improve specificity. Discussion: The ability to detect physiological deterioration before clinically documented complications suggests that unsupervised deep learning approaches can enhance postoperative monitoring, even outside the hospital environment. However, the detection of change points in patients without recorded complications indicates that certain physiological variations may not always be sign of adverse events. This emphasizes the need for improving specificity by incorporating additional clinical parameters, fine tuning the model and refining thresholding methods. Future work will focus on validating the model in a larger patient cohort and optimizing detection algorithms to minimize false alerts
