Time series models show comparable projection performance with joinpoint regression: A comparison using historical cancer data from World Health Organization.

Background: Cancer is one of the major causes of death and the projection of cancer incidences is essential for future healthcare resources planning. Joinpoint regression and average annual percentage change (AAPC) are common approaches for cancer projection, while time series models, traditional ways of trend analysis in statistics, were considered less popular. This study aims to compare these projection methods on seven types of cancers in 31 geographical jurisdictions. Methods: Using data from 66 cancer registries in the World Health Organization, projection models by joinpoint regression, AAPC, and autoregressive integrated moving average with exogenous variables (ARIMAX) were constructed based on 20 years of cancer incidences. The rest of the data upon 20-years of record were used to validate the primary outcomes, namely, 3, 5, and 10-year projections. Weighted averages of mean-square-errors and of percentage errors on predictions were used to quantify the accuracy of the projection results. Results: Among 66 jurisdictions and seven selected cancers, ARIMAX gave the best 5 and 10-year projections for most of the scenarios. When the ten-year projection was concerned, ARIMAX resulted in a mean-square-error (or percentage error) of 2.7% (or 7.2%), compared with 3.3% (or 15.2%) by joinpoint regression and 7.8% (or 15.0%) by AAPC. All the three methods were unable to give reasonable projections for prostate cancer incidence in the US. Conclusion: ARIMAX outperformed the joinpoint regression and AAPC approaches by showing promising accuracy and robustness in projecting cancer incidence rates. In the future, developments in projection models and better applications could promise to improve our ability to understand the trend of disease development, design the intervention strategies, and build proactive public health system.

Machine Learning Clustering for Blood Pressure Variability Applied to Systolic Blood Pressure Intervention Trial (SPRINT) and the Hong Kong Community Cohort.

Visit-to-visit blood pressure variability (BPV) has been shown to be a predictor of cardiovascular disease. We aimed to classify the BPV levels using different machine learning algorithms. Visit-to-visit blood pressure readings were extracted from the SPRINT study in the United States and eHealth cohort in Hong Kong (HK cohort). Patients were clustered into low, medium, and high BPV levels with the traditional quantile clustering and 5 machine learning algorithms including K-means. Clustering methods were assessed by Stability Index. Similarities were assessed by Davies-Bouldin Index and Silhouette Index. Cox proportional hazard regression models were fitted to compare the risk of myocardial infarction, stroke, and heart failure. A total of 8133 participants had average blood pressure measurement 14.7 times in 3.28 years in SPRINT and 1094 participants who had average blood pressure measurement 165.4 times in 1.37 years in HK cohort. Quantile clustering assigned one-third participants as high BPV level, but machine learning methods only assigned 10% to 27%. Quantile clustering is the most stable method (stability index: 0.982 in the SPRINT and 0.948 in the HK cohort) with some levels of clustering similarities (Davies-Bouldin Index: 0.752 and 0.764, respectively). K-means clustering is the most stable across the machine learning algorithms (stability index: 0.975 and 0.911, respectively) with the lowest clustering similarities (Davies-Bouldin Index: 0.653 and 0.680, respectively). One out of 7 in the population was classified with high BPV level, who showed to have higher risk of stroke and heart failure. Machine learning methods can improve BPV classification for better prediction of cardiovascular diseases.

An Integrated Approach of Machine Learning and Systems Thinking for Waiting Time Prediction in an Emergency Department.

OBJECTIVE: The objective of this study is to apply machine learning algorithms for real-time and personalized waiting time prediction in emergency departments. We also aim to introduce the concept of systems thinking to enhance the performance of the prediction models. METHODS: Four popular algorithms were applied: (i) stepwise multiple linear regression; (ii) artificial neural networks; (iii) support vector machines; and (iv) gradient boosting machines. A linear regression model served as a baseline model for comparison. We conducted computational experiments based on a dataset collected from an emergency department in Hong Kong. Model diagnostics were performed, and the results were cross-validated. RESULTS: All the four machine learning algorithms with the use of systems knowledge outperformed the baseline model. The stepwise multiple linear regression reduced the mean-square error by almost 15%. The other three algorithms had similar performances, reducing the mean-square error by approximately 20%. Reductions of 17 - 22% in mean-square error due to the utilization of systems knowledge were observed. DISCUSSION: The multi-dimensional stochasticity arising from the ED environment imposes a great challenge on waiting time prediction. The introduction of the concept of systems thinking led to significant enhancements of the models, suggesting that interdisciplinary efforts could potentially improve prediction performance. CONCLUSION: Machine learning algorithms with the utilization of the systems knowledge could significantly improve the performance of waiting time prediction. Waiting time prediction for less urgent patients is more challenging.

How can we better use Twitter to find a person who got lost due to dementia?

Twitter is a social media platform for online message sharing. The aim of this study is to evaluate the effectiveness of using Twitter to search for people who got lost due to dementia. The online messages on Twitter, i.e., tweets, were collected through an Application Programming Interface. Contents of the tweets were analysed. The personal characteristics, features of tweets and types of Twitter users were collected to investigate their associations with whether a person can be found within a month. Logistic regression was used to identify the features that were useful in finding the missing people. Results showed that the young age of the persons with dementia who got lost, having tweets posted by police departments, and having tweets with photos can increase the chance of being found. Social media is reshaping the human communication pathway, which may lead to future needs on a new patient-care model.