Prediction of Cancer Symptom Trajectory Using Longitudinal Electronic Health Record Data and Long Short-Term Memory Neural Network.

PURPOSE: Ability to predict symptom severity and progression across treatment trajectories would allow clinicians to provide timely intervention and treatment planning. However, such predictions are difficult because of sparse and inconsistent assessment, and simplistic measures such as the last observed symptom severity are often used. The purpose of this study is to develop a model for predicting future cancer symptom experiences on the basis of past symptom experiences. PATIENTS AND METHODS: We performed a retrospective, longitudinal analysis using records of patients with cancer (n = 208) hospitalized between 2008 and 2014. A long short-term memory (LSTM)-based recurrent neural network, a linear regression, and random forest models were trained on previous symptoms experienced and used to predict future symptom trajectories. RESULTS: We found that at least one of three tested models (LSTM, linear regression, and random forest) outperform predictions based solely on the previous clinical observation. LSTM models significantly outperformed linear regression and random forest models in predicting nausea (P < .1) and psychosocial status (P < .01). Linear regression outperformed all models when predicting oral health (P < .01), while random forest outperformed all models when predicting mobility (P < .01) and nutrition (P < .01). CONCLUSION: We can successfully predict patients' symptom trajectories with a prediction model, built with sparse assessment data, using routinely collected nursing documentation. The results of this project can be applied to better individualize symptom management to support cancer patients' quality of life.

Evaluating the performance of machine learning methods for risk estimation of delirium in patients hospitalized from the emergency department.

INTRODUCTION: Delirium is a cerebral dysfunction seen commonly in the acute care setting. It is associated with increased mortality and morbidity and is frequently missed in the emergency department (ED) and inpatient care by clinical gestalt alone. Identifying those at risk of delirium may help prioritize screening and interventions in the hospital setting. OBJECTIVE: Our objective was to leverage electronic health records to identify a clinically valuable risk estimation model for prevalent delirium in patients being transferred from the ED to inpatient units. METHODS: This was a retrospective cohort study to develop and validate a risk model to detect delirium using patient data available from prior visits and ED encounter. Electronic health records were extracted for patients hospitalized from the ED between January 1, 2014, and December 31, 2020. Eligible patients were aged 65 or older, admitted to an inpatient unit from the emergency department, and had at least one DOSS assessment or CAM-ICU recorded within 72 h of hospitalization. Six machine learning models were developed to estimate the risk of delirium using clinical variables including demographic features, physiological measurements, medications administered, lab results, and diagnoses. RESULTS: A total of 28,531 patients met the inclusion criteria with 8057 (28.4%) having a positive delirium screening within the outcome observation period. Machine learning models were compared using the area under the receiver operating curve (AUC). The gradient boosted machine achieved the best performance with an AUC of 0.839 (95% CI, 0.837-0.841). At a 90% sensitivity threshold, this model achieved a specificity of 53.5% (95% CI 53.0%-54.0%) a positive predictive value of 43.5% (95% CI 43.2%-43.9%), and a negative predictive value of 93.1% (95% CI 93.1%-93.2%). A random forest model and L1-penalized logistic regression also demonstrated notable performance with AUCs of 0.837 (95% CI, 0.835-0.838) and 0.831 (95% CI, 0.830-0.833) respectively. CONCLUSION: This study demonstrated the use of machine learning algorithms to identify a combination of variables that enables an estimation of risk of positive delirium screens early in hospitalization to develop prevention or management protocols.

Optimizing warfarin dosing using deep reinforcement learning.

Warfarin is a widely used anticoagulant, and has a narrow therapeutic range. Dosing of warfarin should be individualized, since slight overdosing or underdosing can have catastrophic or even fatal consequences. Despite much research on warfarin dosing, current dosing protocols do not live up to expectations, especially for patients sensitive to warfarin. We propose a deep reinforcement learning-based dosing model for warfarin. To overcome the issue of relatively small sample sizes in dosing trials, we use a Pharmacokinetic/ Pharmacodynamic (PK/PD) model of warfarin to simulate dose-responses of virtual patients. Applying the proposed algorithm on virtual test patients shows that this model outperforms a set of clinically accepted dosing protocols by a wide margin. We tested the robustness of our dosing protocol on a second PK/PD model and showed that its performance is comparable to the set of baseline protocols.

Longitudinal Subgrouping of Patients With Cancer Based on Symptom Experiences: An Integrative Review.

PROBLEM IDENTIFICATION: The purpose of this integrative review is to identify literature describing (a) subgrouping patients with cancer based on symptom experiences and their patterns of symptom changes over time and (b) methodologies of subgrouping patients with cancer based on symptom experiences. LITERATURE SEARCH: PubMed®, CINAHL®, and PsycINFO® were searched through January 2019. DATA EVALUATION: Studies were appraised for patterns of symptom change over time and methodologic approach using the QualSyst quality rating scale. SYNTHESIS: 11 studies met inclusion criteria. Clinical variables that influence symptom patterns were diverse. The most common clustering method was latent variable analysis (73%), and all studies collected symptom data prospectively using survey analysis to assess symptoms. IMPLICATIONS FOR PRACTICE: The majority of studies (91%) observed that the symptom experience within the group of patients with cancer changed over time. Identifying groups of patients with similar symptom experiences is useful to determine which patients need more intensive symptom management over the trajectory of cancer treatment, which is essential to improve symptoms and quality of life.

Machine learning approaches to predict the 1-year-after-initial-AMI survival of elderly patients.

BACKGROUND: While multiple randomized controlled trials (RCTs) are available, their results may not be generalizable to older, unhealthier or less-adherent patients. Observational data can be used to predict outcomes and evaluate treatments; however, exactly which strategy should be used to analyze the outcomes of treatment using observational data is currently unclear. This study aimed to determine the most accurate machine learning technique to predict 1-year-after-initial-acute-myocardial-infarction (AMI) survival of elderly patients and to identify the association of angiotensin-converting- enzyme inhibitors and angiotensin-receptor blockers (ACEi/ARBs) with survival. METHODS: We built a cohort of 124,031 Medicare beneficiaries who experienced an AMI in 2007 or 2008. For analytical purposes, all variables were categorized into nine different groups: ACEi/ARB use, demographics, cardiac events, comorbidities, complications, procedures, medications, insurance, and healthcare utilization. Our outcome of interest was 1-year-post-AMI survival. To solve this classification task, we used lasso logistic regression (LLR) and random forest (RF), and compared their performance depending on category selection, sampling methods, and hyper-parameter selection. Nested 10-fold cross-validation was implemented to obtain an unbiased estimate of performance evaluation. We used the area under the receiver operating curve (AUC) as our primary measure for evaluating the performance of predictive algorithms. RESULTS: LLR consistently showed best AUC results throughout the experiments, closely followed by RF. The best prediction was yielded with LLR based on the combination of demographics, comorbidities, procedures, and utilization. The coefficients from the final LLR model showed that AMI patients with many comorbidities, older ages, or living in a low-income area have a higher risk of mortality 1-year after an AMI. In addition, treating the AMI patients with ACEi/ARBs increases the 1-year-after-initial-AMI survival rate of the patients. CONCLUSIONS: Given the many features we examined, ACEi/ARBs were associated with increased 1-year survival among elderly patients after an AMI. We found LLR to be the best-performing model over RF to predict 1-year survival after an AMI. LLR greatly improved the generalization of the model by feature selection, which implicitly indicates the association between AMI-related variables and survival can be defined by a relatively simple model with a small number of features. Some comorbidities were associated with a greater risk of mortality, such as heart failure and chronic kidney disease, but others were associated with survival such as hypertension, hyperlipidemia, and diabetes. In addition, patients who live in urban areas and areas with large numbers of immigrants have a higher probability of survival. Machine learning methods are helpful to determine outcomes when RCT results are not available.

Modest Clostridiodes difficile infection prediction using machine learning models in a tertiary care hospital.

Previous studies have shown promising results of machine learning (ML) models for predicting health outcomes. We develop and test ML models for predicting Clostridioides difficile infection (CDI) in hospitalized patients. This is a retrospective cohort study conducted during 2015-2017. All inpatients tested for C. difficile were included. CDI was defined as having a positive glutamate dehydrogenase and toxin results. We restricted analyses to the first record of C. difficile testing per patient. Of 3514 patients tested, 136 (4%) had CDI. Age and antibiotic use within 90 days before C. difficile testing were associated with CDI (P < 0.01). We tested 10 ML methods with and without resampling. Logistic regression, random forest and naïve Bayes models yielded the highest AUC ROC performance: 0.6. Predicting CDI was difficult in our cohort of patients tested for CDI. Multiple ML models yielded only modest results in a real-world population of hospitalized patients tested for CDI.

Modeling and maximizing influence diffusion in social networks for viral marketing.

Modeling influence diffusion in social networks is an important challenge. We investigate influence-diffusion modeling and maximization in the setting of viral marketing, in which a node's influence is measured by the number of nodes it can activate to adopt a new technology or purchase a new product. One of the fundamental problems in viral marketing is to find a small set of initial adopters who can trigger the most further adoptions through word-of-mouth-based influence propagation in the network. We propose a novel multiple-path asynchronous threshold (MAT) model, in which we quantify influence and track its diffusion and aggregation. Our MAT model captures not only direct influence from neighboring influencers but also indirect influence passed along by messengers. Moreover, our MAT framework models influence attenuation along diffusion paths, temporal influence decay, and individual diffusion dynamics. Our work is an important step toward a more realistic diffusion model. Further, we develop an effective and efficient heuristic to tackle the influence-maximization problem. Our experiments on four real-life networks demonstrate its excellent performance in terms of both influence spread and time efficiency. Our work provides preliminary but significant insights and implications for diffusion research and marketing practice.

Adverse drug effect detection.

Large collections of electronic patient records provide abundant but under-explored information on the real-world use of medicines. Although they are maintained for patient administration, they provide a broad range of clinical information for data analysis. One growing interest is drug safety signal detection from these longitudinal observational data. In this paper, we proposed two novel algorithms-a likelihood ratio model and a Bayesian network model-for adverse drug effect discovery. Although the performance of these two algorithms is comparable to the state-of-the-art algorithm, Bayesian confidence propagation neural network, the combination of three works better due to their diversity in solutions. Since the actual adverse drug effects on a given dataset cannot be absolutely determined, we make use of the simulated observational medical outcomes partnership (OMOP) dataset constructed with the predefined adverse drug effects to evaluate our methods. Experimental results show the usefulness of the proposed pattern discovery method on the simulated OMOP dataset by improving the standard baseline algorithm-chi-square-by 23.83%.

Building a computer program to support children, parents, and distraction during healthcare procedures.

This secondary data analysis used data mining methods to develop predictive models of child risk for distress during a healthcare procedure. Data used came from a study that predicted factors associated with children's responses to an intravenous catheter insertion while parents provided distraction coaching. From the 255 items used in the primary study, 44 predictive items were identified through automatic feature selection and used to build support vector machine regression models. Models were validated using multiple cross-validation tests and by comparing variables identified as explanatory in the traditional versus support vector machine regression. Rule-based approaches were applied to the model outputs to identify overall risk for distress. A decision tree was then applied to evidence-based instructions for tailoring distraction to characteristics and preferences of the parent and child. The resulting decision support computer application, titled Children, Parents and Distraction, is being used in research. Future use will support practitioners in deciding the level and type of distraction intervention needed by a child undergoing a healthcare procedure.

A decision support system for cost-effective diagnosis.

OBJECTIVE: Speed, cost, and accuracy are three important goals in disease diagnosis. This paper proposes a machine learning-based expert system algorithm to optimize these goals and assist diagnostic decisions in a sequential decision-making setting. METHODS: The algorithm consists of three components that work together to identify the sequence of diagnostic tests that attains the treatment or no test threshold probability for a query case with adequate certainty: lazy-learning classifiers, confident diagnosis, and locally sequential feature selection (LSFS). Speed-based and cost-based objective functions can be used as criteria to select tests. RESULTS: Results of four different datasets are consistent. All LSFS functions significantly reduce tests and costs. Average cost savings for heart disease, thyroid disease, diabetes, and hepatitis datasets are 50%, 57%, 22%, and 34%, respectively. Average test savings are 55%, 73%, 24%, and 39%, respectively. Accuracies are similar to or better than the baseline (the classifier that uses all available tests in the dataset). CONCLUSION: We have demonstrated a new approach that dynamically estimates and determines the optimal sequence of tests that provides the most information (or disease probability) based on a patient's available information.

The optimal diagnostic decision sequence.

We describe a data mining model for constructing an optimal diagnostic sequence that assists cost-effective sequential decisions. We use heuristic search, i.e., hill climbing and genetic algorithms (GAs), and the evaluation function of cost-based Mean Accuracy Gain (cMAG), which is provided by SVM classifiers, to find this optimal sequence. GA can find a good sequence because of the ability to escape from local optima.

Building a hospital referral expert system with a Prediction and Optimization-Based Decision Support System algorithm.

This study presents a new method for constructing an expert system using a hospital referral problem as an example. Many factors, such as institutional characteristics, patient risks, traveling distance, and chances of survival and complications should be included in the hospital-selection decision. Ideally, each patient should be treated individually, with the decision process including not only their condition but also their beliefs about trade-offs among the desired hospital features. An expert system can help with this complex decision, especially when numerous factors are to be considered. We propose a new method, called the Prediction and Optimization-Based Decision Support System (PODSS) algorithm, which constructs an expert system without an explicit knowledge base. The algorithm obtains knowledge on its own by building machine learning classifiers from a collection of labeled cases. In response to a query, the algorithm gives a customized recommendation, using an optimization step to help the patient maximize the probability of achieving a desired outcome. In this case, the recommended hospital is the optimal solution that maximizes the probability of the desired outcome. With proper formulation, this expert system can combine multiple factors to give hospital-selection decision support at the individual level.

A data mining technique for risk-stratification diagnosis.

We describe a data mining model for sequential diagnosis, called the Optimal Decision Path Finder (ODPF), which is built based on the idea of risk stratification. A filter was used to stratify patients depending on ease of diagnosis, and a series of patient-specific classifiers was built to diagnose with confidence while reducing exam cost. Results show that applying stratification to data mining can improve the diagnostic performance and reduce waste of medical resource. This resulting model can assist the physician in triage decisions.

Application of artificial neural network-based survival analysis on two breast cancer datasets.

This paper applies artificial neural networks (ANNs) to the survival analysis problem. Because ANNs can easily consider variable interactions and create a non-linear prediction model, they offer more flexible prediction of survival time than traditional methods. This study compares ANN results on two different breast cancer datasets, both of which use nuclear morphometric features. The results show that ANNs can successfully predict recurrence probability and separate patients with good (more than five years) and bad (less than five years) prognoses. Results are not as clear when the separation is done within subgroups such as lymph node positive or negative.

Predicting outcomes of hospitalization for heart failure using logistic regression and knowledge discovery methods.

The purpose of this study is to determine the best prediction of heart failure outcomes, resulting from two methods -- standard epidemiologic analysis with logistic regression and knowledge discovery with supervised learning/data mining. Heart failure was chosen for this study as it exhibits higher prevalence and cost of treatment than most other hospitalized diseases. The prevalence of heart failure has exceeded 4 million cases in the U.S.. Findings of this study should be useful for the design of quality improvement initiatives, as particular aspects of patient comorbidity and treatment are found to be associated with mortality. This is also a proof of concept study, considering the feasibility of emerging health informatics methods of data mining in conjunction with or in lieu of traditional logistic regression methods of prediction. Findings may also support the design of decision support systems and quality improvement programming for other diseases.

Computer-generated nuclear features compared with axillary lymph node status and tumor size as indicators of breast cancer survival.

The extent to which malignant cells deviate from normal is generally accepted to be a prognostic indicator. However, assessing the degree of deviation has been subjective and poorly reproducible. Our goal is to develop a computer program for objectively measuring nuclear size, shape, and texture from histologic slides and to make the program available on the Internet. We used this program to analyze 353 histologic sections obtained from patients with invasive breast cancer who were diagnosed and treated from 1981 through 1995 and who had determinable outcomes. The median follow-up was 8.3 years. We compared the relationship of survival with our computer-derived nuclear features versus axillary lymph node status and tumor size. We believe that our results are generally applicable because our patient survival, when stratified by lymph node status, was similar to that of the 24,000 breast cancer patients in the National Cancer Institute's Surveillance, Epidemiology, and End Results program. In multivariate analysis, the strongest prognostic factor was the largest nuclear area, followed by tumor size and the extent of axillary lymph node involvement. The mean area of the 3 largest nuclei when combined with tumor size identified 30% of all breast cancer patients who had an 87% 15-year breast cancer-specific survival. Inclusion of lymph node status added little to this 2-factor model. Routine axillary lymph node surgery for prognostic purposes may become unnecessary, because nuclear features may provide sufficient information.