Mustafa Khanbhai1, Leigh Warren2, Joshua Symons3, Kelsey Flott2, Stephanie Harrison-White4, Dave Manton2, Ara Darzi2, Erik Mayer2. 1. Centre for Health Policy, Institute of Global Health Innovation, Imperial College London, QEQM, St Mary's Hospital, W2 1NY, UK. Electronic address: m.khanbhai@imperial.ac.uk. 2. Patient Safety Translational Research Centre, Institute of Global Health Innovation, Imperial College London, London, UK. 3. Big Data and Analytical Unit, Institute of Global Health Innovation, Imperial College London, London, UK. 4. Patient Experience and Improvement, Imperial College NHS Healthcare Trust, London, UK.
Abstract
BACKGROUND: Patient centred care necessitates that healthcare experiences and perceived outcomes be considered across all transitions of care. Information encoded within free-text patient experience comments relating to transitions of care are not captured in a systematic way due to the manual resource required. We demonstrate the use of natural language processing (NLP) to extract meaningful information from the Friends and Family Test (FFT). METHODS: Free-text fields identifying favourable service ("What did we do well?") and areas requiring improvement ("What could we do better?") were extracted from 69,285 FFT reports across four care settings at a secondary care National Health Service (NHS) hospital. Sentiment and patient experience themes were coded by three independent coders to produce a training dataset. The textual data was standardised with a series of pre-processing techniques and the performance of six machine learning (ML) models was obtained. The best performing ML model was applied to predict the themes and sentiment from the remaining reports. Comments relating to transitions of care were extracted, categorised by sentiment, and care setting to identify the most frequent words/combinations presented as tri-grams and word clouds. RESULTS: The support vector machine (SVM) ML model produced the highest accuracy in predicting themes and sentiment. The most frequent single words relating to transition and continuity with a negative sentiment were "discharge" in inpatients and Accident and Emergency, "appointment" in outpatients, and "home' in maternity. Tri-grams identified from the negative sentiments such as 'seeing different doctor', 'information aftercare lacking', 'improve discharge process' and 'timing discharge letter' have highlighted some of the problems with care transitions. None of this information was available from the quantitative data. CONCLUSIONS: NLP can be used to identify themes and sentiment from patient experience survey comments relating to transitions of care in all four healthcare settings. With the help of a quality improvement framework, findings from our analysis may be used to guide patient-centred interventions to improve transitional care processes.
BACKGROUND: Patient centred care necessitates that healthcare experiences and perceived outcomes be considered across all transitions of care. Information encoded within free-text patient experience comments relating to transitions of care are not captured in a systematic way due to the manual resource required. We demonstrate the use of natural language processing (NLP) to extract meaningful information from the Friends and Family Test (FFT). METHODS: Free-text fields identifying favourable service ("What did we do well?") and areas requiring improvement ("What could we do better?") were extracted from 69,285 FFT reports across four care settings at a secondary care National Health Service (NHS) hospital. Sentiment and patient experience themes were coded by three independent coders to produce a training dataset. The textual data was standardised with a series of pre-processing techniques and the performance of six machine learning (ML) models was obtained. The best performing ML model was applied to predict the themes and sentiment from the remaining reports. Comments relating to transitions of care were extracted, categorised by sentiment, and care setting to identify the most frequent words/combinations presented as tri-grams and word clouds. RESULTS: The support vector machine (SVM) ML model produced the highest accuracy in predicting themes and sentiment. The most frequent single words relating to transition and continuity with a negative sentiment were "discharge" in inpatients and Accident and Emergency, "appointment" in outpatients, and "home' in maternity. Tri-grams identified from the negative sentiments such as 'seeing different doctor', 'information aftercare lacking', 'improve discharge process' and 'timing discharge letter' have highlighted some of the problems with care transitions. None of this information was available from the quantitative data. CONCLUSIONS: NLP can be used to identify themes and sentiment from patient experience survey comments relating to transitions of care in all four healthcare settings. With the help of a quality improvement framework, findings from our analysis may be used to guide patient-centred interventions to improve transitional care processes.
Authors: Marieke M van Buchem; Olaf M Neve; Ilse M J Kant; Ewout W Steyerberg; Hileen Boosman; Erik F Hensen Journal: BMC Med Inform Decis Mak Date: 2022-07-15 Impact factor: 3.298