[Paper Review] Smart City Analytics: Ensemble-Learned Prediction of Citizen Home Care

Casper Worm Hansen, Christian Hansen+2 more|arXiv (Cornell University)|Aug 21, 2017

Human Mobility and Location-Based Analysis5 references3 citations

TL;DR

This paper proposes an ensemble learning approach to predict large increases in home care hours for citizens in Copenhagen, using either linear (logistic regression) or non-linear (random forests) models trained on historical data. The method achieves a state-of-the-art AUC of 0.715, demonstrating strong performance even with limited training data, and identifies recent care patterns as the most predictive features.

ABSTRACT

We present an ensemble learning method that predicts large increases in the hours of home care received by citizens. The method is supervised, and uses different ensembles of either linear (logistic regression) or non-linear (random forests) classifiers. Experiments with data available from 2013 to 2017 for every citizen in Copenhagen receiving home care (27,775 citizens) show that prediction can achieve state of the art performance as reported in similar health related domains (AUC=0.715). We further find that competitive results can be obtained by using limited information for training, which is very useful when full records are not accessible or available. Smart city analytics does not necessarily require full city records. To our knowledge this preliminary study is the first to predict large increases in home care for smart city analytics.

Motivation & Objective

• To develop a predictive model for large increases in citizen home care hours using smart city data.
• To evaluate the performance of ensemble learning methods in forecasting long-term home care needs.
• To assess the impact of varying data availability levels on model accuracy.
• To identify the most predictive features for home care escalation.
• To demonstrate the feasibility of using limited, low-cost data for effective smart city health analytics.

Proposed method

• The method frames home care prediction as a binary classification problem: predicting whether a citizen will experience a 6+ hour increase in home care over the next 3 months.
• It employs a two-level ensemble learning architecture: level 0 models (logistic regression or random forests) are trained on sequential data chunks, and their predictions are combined via a level 1 meta-model.
• Level 0 models are retrained monthly using either the previous month’s data (IL1) or all prior data (IL2), forming a dynamic ensemble.
• The final prediction is generated by combining level 0 model outputs into a meta-instance vector, which is then classified by a level 1 model.
• Regularized logistic regression and random forest classifiers are used as base learners, with hyperparameters tuned for optimal performance.
• Performance is evaluated using AUC, with ablation studies across five information levels (IL1–IL4) to assess data efficiency.

Experimental results

Research questions

• RQ1 Can ensemble learning effectively predict large increases in home care hours for individual citizens in a smart city context?
• RQ2 How does model performance vary when training data is limited to only recent or minimal information?
• RQ3 Which features are most predictive of a significant increase in home care needs?
• RQ4 How does the proposed method compare to heuristic baselines based on historical patterns?
• RQ5 To what extent can binary indicators of service receipt replace detailed clinical data for accurate prediction?

Key findings

• The ensemble model achieved an AUC of 0.715 when trained on all available historical data, outperforming both heuristic baselines (AUC 0.548 and 0.634).
• The best-performing model used all prior data for training, indicating that long-term historical patterns are critical for accurate prediction.
• The number of prior large increases in home care over the last 3 months was the most predictive feature, with a weight of 0.38.
• Other key features included hospitalization (0.22), sick care receipt (0.21), weekend home care (0.21), and age (0.14).
• Using only binary indicators of service receipt (IL2b) yielded performance comparable to full data, suggesting low-cost data can be highly effective.
• The model’s performance remained stable across information levels, with AUC consistently above 0.710 when using comprehensive data, indicating robustness to data sparsity.