Improving Prediction of Risk of Hospital Admission in Chronic Obstructive Pulmonary Disease: Application of Machine Learning to Telemonitoring Data.

Background: Telemonitoring of symptoms and physiological signs has been suggested as a means of early detection of chronic obstructive pulmonary disease (COPD) exacerbations, with a view to instituting timely treatment. However, algorithms to identify exacerbations result in frequent false-positive results and increased workload. Machine learning, when applied to predictive modelling, can determine patterns of risk factors useful for improving prediction quality. Objective: Our objectives were to (1) establish whether machine learning techniques applied to telemonitoring datasets improve prediction of hospital admissions and decisions to start corticosteroids, and (2) determine whether the addition of weather data further improves such predictions. Methods: We used daily symptoms, physiological measures, and medication data, with baseline demography, COPD severity, quality of life, and hospital admissions from a pilot and large randomized controlled trial of telemonitoring in COPD. We linked weather data from the United Kingdom meteorological service. We used feature selection and extraction techniques for time series to construct up to 153 predictive patterns (features) from symptom, medication, and physiological measurements. We used the resulting variables to construct predictive models fitted to training sets of patients and compared them with common symptom-counting algorithms. Results: We had a mean 363 days of telemonitoring data from 135 patients. The two most practical traditional score-counting algorithms, restricted to cases with complete data, resulted in area under the receiver operating characteristic curve (AUC) estimates of 0.60 (95% CI 0.51-0.69) and 0.58 (95% CI 0.50-0.67) for predicting admissions based on a single day's readings. However, in a real-world scenario allowing for missing data, with greater numbers of patient daily data and hospitalizations (N=57,150, N+=55, respectively), the performance of all the traditional algorithms fell, including those based on 2 days' data. One of the most frequently used algorithms performed no better than chance. All considered machine learning models demonstrated significant improvements; the best machine learning algorithm based on 57,150 episodes resulted in an aggregated AUC of 0.74 (95% CI 0.67-0.80). Adding weather data measurements did not improve the predictive performance of the best model (AUC 0.74, 95% CI 0.69-0.79). To achieve an 80% true-positive rate (sensitivity), the traditional algorithms were associated with an 80% false-positive rate: our algorithm halved this rate to approximately 40% (specificity approximately 60%). The machine learning algorithm was moderately superior to the best symptom-counting algorithm (AUC 0.77, 95% CI 0.74-0.79 vs AUC 0.66, 95% CI 0.63-0.68) at predicting the need for corticosteroids. Conclusions: Early detection and management of COPD remains an important goal given its huge personal and economic costs. Machine learning approaches, which can be tailored to an individual's baseline profile and can learn from experience of the individual patient, are superior to existing predictive algorithms and show promise in achieving this goal.