TL;DR
This paper introduces a novel machine learning approach for the early detection of Urinary Tract Infections (UTIs) in individuals with dementia, leveraging passive in-home sensor data. By analyzing activity and physiological patterns, the system aims to provide timely alerts to clinicians, potentially reducing hospitalizations and improving patient outcomes.
What Problem Does This Paper Address?
Urinary Tract Infections (UTIs) are a common and serious health issue, particularly for older adults and especially for those living with dementia (PLWD). A major challenge in this population is the difficulty in reporting symptoms due to cognitive impairment, leading to delayed diagnosis and potentially severe complications like sepsis and hospital admissions. Current diagnostic methods, primarily urine sample analysis, are reactive and can be time-consuming. There's a clear need for proactive, non-intrusive methods to identify UTIs early, before they escalate.
Methods
The core of this research is a machine learning pipeline designed to predict UTI risk. The authors explored several classification models, ultimately selecting Logistic Regression (LR) as the best performer when trained on 3 consecutive days of data.
Key methodological aspects include:
- Feature Engineering: They engineered features from passively collected data that represent known UTI symptoms. This included analyzing bathroom visit frequency, sleep patterns (sleep stages, time awake), nocturnal respiratory rate, and nocturnal heart rate.
- Passive Monitoring: The system relies on low-cost, off-the-shelf sensors (PIR motion sensors and sleep mats) to collect data unobtrusively in the home environment.
- Explainable AI (SHAP): To understand why the model makes certain predictions, they employed SHapley Additive exPlanations (SHAP) to identify the most influential features.
- Risk Stratification: To make the model clinically actionable, they developed a risk stratification system that categorizes predictions into "Green" (minimal risk), "Amber" (medium risk), and "Red" (high risk) alerts. This allows for balancing sensitivity and specificity based on clinical needs and resources.
- Validation Strategies: The model was evaluated using two distinct approaches:
- "Date Split": Temporal splitting of data (80% train, 20% test) to simulate real-world prediction on future data from the same cohort.
- "Date-ID Split": Leave-one-out cross-validation on participants to ensure the model generalizes to individuals not seen during training.
Datasets
The study utilized a dataset of 27,828 person-days from 117 participants living with dementia. This data was collected between June 2021 and December 2022. The dataset includes:
- Passive Sensor Data: Activity (PIR sensors) and physiological data (sleep mat measuring breathing rate, heart rate, and sleep states).
- Clinician-Labeled UTI Events: Urine samples were collected and analyzed, with clinicians determining UTI status. This resulted in 258 labeled urine samples from 64 participants, with 81 confirmed positive UTI cases. To extend the labeled data, preceding and succeeding days were also labeled, resulting in 1752 labeled days (534 positive, 1218 negative).
Key Insights and Ideas
- Passive Sensing is Viable for UTI Detection: The study demonstrates that continuous, passive monitoring of everyday activities and physiological signals can capture subtle changes indicative of UTIs, even in individuals who struggle to communicate their symptoms.
- Bathroom Behavior is a Strong Predictor: Increased bathroom visit frequency, particularly at night, emerged as a significant indicator of UTI risk. This aligns with common UTI symptoms like increased urination frequency and urgency.
- Nocturnal Physiology Matters: Disturbances in sleep patterns, such as increased time awake at night and variations in respiratory and heart rates during sleep, are strongly associated with higher UTI risk. This suggests that UTIs disrupt sleep quality and physiological regulation.
- Explainability Builds Trust: The use of SHAP values allows clinicians to understand which specific data points are driving a UTI risk prediction, fostering trust and enabling more informed clinical decisions.
- Risk Stratification Enhances Clinical Utility: The ability to stratify risk into actionable alert levels (Green, Amber, Red) is crucial for practical implementation, allowing healthcare providers to manage alerts effectively and tailor interventions.
Key Results
- Model Performance (Date-ID Split, after stratification):
- Sensitivity: 74.7% (95% CI = 67.9–81.5)
- Specificity: 87.9% (95% CI = 85.0–90.9)
- Model Performance (Date Split, after stratification):
- Sensitivity: 69.0% (95% CI = 64.4–73.5)
- Specificity: 94.1% (95% CI = 92.0–96.2)
- Early Detection: The model showed promise in predicting UTIs up to 4 days prior to the recorded clinical diagnosis, with a sensitivity of 64.4% and specificity of 71.9% at 4 days prior.
- Top Predictive Features: Number of previous UTI events, standard deviation of nocturnal respiratory rate, nocturnal average heart rate, and number of nocturnal awake states were identified as key predictors.
Limitations and Future Work
- Distinguishing Individuals: The current PIR sensors cannot differentiate between the study participant and other household members, which could lead to false positives. Future work could explore methods for passively collecting personalized data.
- Data Imbalance: The study acknowledges that the model tends to overestimate UTI risk, likely due to data imbalance in the training set. While risk stratification helps, further work is needed to calibrate the risk groups more precisely.
- Generalizability: While the study focused on PLWD, applying these techniques to a broader cohort, including older adults in community living environments (care homes, assisted living), is suggested.
- Clinical Outcomes: Future research should investigate the impact of this monitoring system on actual clinical outcomes, such as hospital admission rates and quality of life.
- Continuous Improvement: The authors plan to continually train the model on new data and monitor its performance and feature importance to ensure ongoing accuracy and clinical relevance.
Summary updated