Using random forests to model 90-day hometime in people with stroke

The association between inpatient rehabilitation intensity and outcomes after stroke in Ontario, Canada

BACKGROUND

Several studies have demonstrated improved outcomes poststroke when higher intensity rehabilitation is provided. Canadian Stroke Best Practice Recommendations advise patients receive 180 min of therapy time per day; however, the exact amount required to reach benefit is unknown.

AIMS

The primary aim of this study was to determine the association between rehabilitation intensity (RI) and total Functional Independence Measure (FIM) Instrument change. Secondary aims included determining the association between RI and discharge location, 90-day home time, rehabilitation effectiveness, and motor and cognitive FIM change.

METHODS

A retrospective cohort study was conducted using available administrative databases of acute stroke patients discharged to inpatient rehabilitation facilities in Ontario, Canada, from January 2017 to December 2021. RI was defined as number of minutes per day of direct therapy by all providers divided by rehabilitation length of stay. The association between RI and the outcomes of interest were analyzed using regression models with restricted cubic splines.

RESULTS

A total of 12,770 individuals were included. Mean age of the sample was 72.6 years, 46.0% of individuals were female, and 87.6% had an ischemic stroke. Mean RI was 74.7 min (range: 5-162 min) per day. Increased RI was associated with an increase in mean FIM change. However, there was diminishing incremental increase after reaching 95 min/day. Increased RI was positively associated with motor and cognitive FIM change, rehabilitation effectiveness, 90-day home time, and discharge to preadmission setting. Higher RI was associated with a lower likelihood of discharge to long-term care.

CONCLUSIONS

None of the patients met the recommended RI of 180 min/day based on the Canadian Stroke Best Practice Recommendations. Despite this, higher intensity was associated with better outcomes. Given that most positive associations were observed with a RI ⩾95 min/day, this may be a more feasible target.

How individuals' opinions influence society's resistance to epidemics: an agent-based model approach

BACKGROUND

Protecting public health from infectious diseases often relies on the cooperation of citizens, especially when self-care interventions are the only viable tools for disease mitigation. Accordingly, social aspects related to public opinion have been studied in the context of the recent COVID-19 pandemic. However, a comprehensive understanding of the effects of opinion-related factors on disease spread still requires further exploration.

METHODS

We propose an agent-based simulation framework incorporating opinion dynamics within an epidemic model based on the assumption that mass media channels play a leading role in opinion dynamics. The model simulates how opinions about preventive interventions change over time and how these changes affect the cumulative number of cases. We calibrated our simulation model using YouGov survey data and WHO COVID-19 new cases data from 15 different countries. Based on the calibrated models, we examine how different opinion-related factors change the consequences of the epidemic. We track the number of total new infections for analysis.

RESULTS

Our results reveal that the initial level of public opinion on preventive interventions has the greatest impact on the cumulative number of cases. Its normalized permutation importance varies between 69.67% and 96.65% in 15 models. The patterns shown in the partial dependence plots indicate that other factors, such as the usage of the pro-intervention channel and the response time of media channels, can also bring about substantial changes in disease dynamics, but only within specific ranges of the dominant factor.

CONCLUSIONS

Our results reveal the importance of public opinion on intervention during the early stage of the pandemic in protecting public health. The findings suggest that persuading the public to take actions they may be hesitant about in the early stages of epidemics is very costly because taking early action is critical for mitigating infectious diseases. Other opinion-related factors can also lead to significant changes in epidemics, depending on the average level of public opinion in the initial stage. These findings underscore the importance of media channels and authorities in delivering accurate information and persuading community members to cooperate with public health policies.

Predicting 90-day prognosis for patients with stroke: a machine learning approach

Background

Stroke is a significant global health burden and ranks as the second leading cause of death worldwide.

Objective

This study aims to develop and evaluate a machine learning-based predictive tool for forecasting the 90-day prognosis of stroke patients after discharge as measured by the modified Rankin Score.

Methods

The study utilized data from a large national multiethnic stroke registry comprising 15,859 adult patients diagnosed with ischemic or hemorrhagic stroke. Of these, 7,452 patients satisfied the study's inclusion criteria. Feature selection was performed using the correlation and permutation importance methods. Six classifiers, including Random Forest (RF), Classification and Regression Tree, Linear Discriminant Analysis, Support Vector Machine, and k-Nearest Neighbors, were employed for prediction.

Results

The RF model demonstrated superior performance, achieving the highest accuracy (0.823) and excellent discrimination power (AUC 0.893). Notably, stroke type, hospital acquired infections, admission location, and hospital length of stay emerged as the top-ranked predictors.

Conclusion

The RF model shows promise in predicting stroke prognosis, enabling personalized care plans and enhanced preventive measures for stroke patients. Prospective validation is essential to assess its real-world clinical performance and ensure successful implementation across diverse healthcare settings.

Development and evaluation of machine learning models and nomogram for the prediction of severe acute pancreatitis

BACKGROUND AND AIM

Severe acute pancreatitis (SAP) in patients progresses rapidly and can cause multiple organ failures associated with high mortality. We aimed to train a machine learning (ML) model and establish a nomogram that could identify SAP, early in the course of acute pancreatitis (AP).

METHODS

In this retrospective study, 631 patients with AP were enrolled in the training cohort. For predicting SAP early, five supervised ML models were employed, such as random forest (RF), K-nearest neighbors (KNN), and naive Bayes (NB), which were evaluated by accuracy (ACC) and the areas under the receiver operating characteristic curve (AUC). The nomogram was established, and the predictive ability was assessed by the calibration curve and AUC. They were externally validated by an independent cohort of 109 patients with AP.

RESULTS

In the training cohort, the AUC of RF, KNN, and NB models were 0.969, 0.954, and 0.951, respectively, while the AUC of the Bedside Index for Severity in Acute Pancreatitis (BISAP), Ranson and Glasgow scores were only 0.796, 0.847, and 0.837, respectively. In the validation cohort, the RF model also showed the highest AUC, which was 0.961. The AUC for the nomogram was 0.888 and 0.955 in the training and validation cohort, respectively.

CONCLUSIONS

Our findings suggested that the RF model exhibited the best predictive performance, and the nomogram provided a visual scoring model for clinical practice. Our models may serve as practical tools for facilitating personalized treatment options and improving clinical outcomes through pre-treatment stratification of patients with AP.

Predicting the behavioral intentions of hospice and palliative care providers from real-world data using supervised learning: A cross-sectional survey study

Background

Hospice and palliative care (HPC) aims to improve end-of-life quality and has received much more attention through the lens of an aging population in the midst of the coronavirus disease pandemic. However, several barriers remain in China due to a lack of professional HPC providers with positive behavioral intentions. Therefore, we conducted an original study introducing machine learning to explore individual behavioral intentions and detect factors of enablers of, and barriers to, excavating potential human resources and improving HPC accessibility.

Methods

A cross-sectional study was designed to investigate healthcare providers' behavioral intentions, knowledge, attitudes, and practices in hospice care (KAPHC) with an indigenized KAPHC scale. Binary Logistic Regression and Random Forest Classifier (RFC) were performed to model impacting and predict individual behavioral intentions.

Results

The RFC showed high sensitivity (accuracy = 0.75; F1 score = 0.84; recall = 0.94). Attitude could directly or indirectly improve work enthusiasm and is the most efficient approach to reveal behavioral intentions. Continuous practice could also improve individual confidence and willingness to provide HPC. In addition, scientific knowledge and related skills were the foundation of implementing HPC.

Conclusion

Individual behavioral intention is crucial for improving HPC accessibility, particularly at the initial stage. A well-trained RFC can help estimate individual behavioral intentions to organize a productive team and promote additional policies.

External validation of the Passive Surveillance Stroke Severity Indicator

BACKGROUND

The Passive Surveillance Stroke Severity (PaSSV) Indicator was derived to estimate stroke severity from variables in administrative datasets but has not been externally validated.

METHODS

We used linked administrative datasets to identify patients with first hospitalization for acute stroke between 2007-2018 in Alberta, Canada. We used the PaSSV indicator to estimate stroke severity. We used Cox proportional hazard models and evaluated the change in hazard ratios and model discrimination for 30-day and 1-year case fatality with and without PaSSV. Similar comparisons were made for 90-day home time thresholds using logistic regression. We also linked with a clinical registry to obtain National Institutes of Health Stroke Scale (NIHSS) and compared estimates from models without stroke severity, with PaSSV, and with NIHSS.

RESULTS

There were 28,672 patients with acute stroke in the full sample. In comparison to no stroke severity, addition of PaSSV to the 30-day case fatality models resulted in improvement in model discrimination (C-statistic 0.72 [95%CI 0.71-0.73] to 0.80 [0.79-0.80]). After adjustment for PaSSV, admission to a comprehensive stroke center was associated with lower 30-day case fatality (adjusted hazard ratio changed from 1.03 [0.96-1.10] to 0.72 [0.67-0.77]). In the registry sample (N = 1328), model discrimination for 30-day case fatality improved with the inclusion of stroke severity. Results were similar for 1-year case fatality and home time outcomes.

CONCLUSION

Addition of PaSSV improved model discrimination for case fatality and home time outcomes. The validity of PASSV in two Canadian provinces suggests that it is a useful tool for baseline risk adjustment in acute stroke.

Comparing regression modeling strategies for predicting hometime

Background

Hometime, the total number of days a person is living in the community (not in a healthcare institution) in a defined period of time after a hospitalization, is a patient-centred outcome metric increasingly used in healthcare research. Hometime exhibits several properties which make its statistical analysis difficult: it has a highly non-normal distribution, excess zeros, and is bounded by both a lower and upper limit. The optimal methodology for the analysis of hometime is currently unknown.

Methods

Using administrative data we identified adult patients diagnosed with stroke between April 1, 2010 and December 31, 2017 in Ontario, Canada. 90-day hometime and clinically relevant covariates were determined through administrative data linkage. Fifteen different statistical and machine learning models were fit to the data using a derivation sample. The models’ predictive accuracy and bias were assessed using an independent validation sample.

Results

Seventy-five thousand four hundred seventy-five patients were identified (divided into a derivation set of 49,402 and a test set of 26,073). In general, the machine learning models had lower root mean square error and mean absolute error than the statistical models. However, some statistical models resulted in lower (or equal) bias than the machine learning models. Most of the machine learning models constrained predicted values between the minimum and maximum observable hometime values but this was not the case for the statistical models. The machine learning models also allowed for the display of complex non-linear interactions between covariates and hometime. No model captured the non-normal bucket shaped hometime distribution.

Conclusions

Overall, no model clearly outperformed the others. However, it was evident that machine learning methods performed better than traditional statistical methods. Among the machine learning methods, generalized boosting machines using the Poisson distribution as well as random forests regression were the best performing. No model was able to capture the bucket shaped hometime distribution and future research on factors which are associated with extreme values of hometime that are not available in administrative data is warranted.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12874-021-01331-9.