S Khalid

A Two-Class Approach to the Detection of Physiological Deterioration in Patient Vital Signs, With Clinical Label Refinement

Hospital patient outcomes can be improved by the early identification of physiological deterioration. Automatic methods of detecting patient deterioration in vital-sign data typically attempt to identify deviations from assumed “normal” physiological conditions, which is a one-class approach to classification. This paper investigates the use of a two-class approach, in which “abnormal” physiology is modeled explicitly. The success of such a method relies on the accuracy of data labels provided by clinical experts, which may be incomplete (due to large dataset size) or imprecise (due to clinical labels covering intervals, rather than each data point within those intervals). We propose a novel method of refining clinical labels such that the two-class classification approach may be adopted for identifying patient deterioration. We demonstrate the effectiveness of the proposed methods using a large dataset acquired in a 24-bed hospital step-down unit.

An Unsupervised Learning Model for Pattern Recognition in Routinely Collected Healthcare Data

This study examines a large routinely collected healthcare database containing patient-level self-reported outcomes following knee replacement surgery. A model based on unsupervised machine learning methods, including k-means and hierarchical clustering, is proposed to detect patterns of pain experienced by patients and to derive subgroups of patients with different outcomes based on their pain characteristics. Results showed the presence of between two and four different sub-groups of patients based on their pain characteristics. Challenges associated with unsupervised learning using real-world data are described and an approach for evaluating models in the presence of unlabelled data using internal and external cluster evaluation techniques is presented, that can be extended to other unsupervised learning applications within healthcare and beyond. To our knowledge, this is the first study proposing an unsupervised learning model for characterising painbased patient subgroups using the UK NHS PROMs database.

Risk of venous thromboembolism among users of different anti-osteoporosis drugs: a population-based cohort analysis including over 200,000 participants from Spain and the UK.

SummaryThe venous thromboembolism risk among anti-osteoporotics is unknown. In this primary care study, the risk with other bisphosphonates [1.05 (0.94–1.18) and 0.96 (0.78–1.18)], strontium [0.90 (0.61–1.34) and 1.19 (0.82–1.74)], in the UK and Spain respectively, and denosumab [1.77 (0.25–12.66)] and teriparatide [1.27 (0.59–2.71)] in Spain, did not differ versus alendronate.IntroductionMost of the known adverse drug reactions described for anti-osteoporosis medication (AOM) have been described in studies comparing AOM users to non-users. We aimed to compare the risk of venous thromboembolism (VTE) among incident users of different AOM compared to alendronate (first line therapy).MethodsTwo cohort studies were performed using data from the UK (CPRD) and Spain (BIFAP) primary care records separately. All patients aged ≥u200950xa0years with at least 1xa0year of data available and a new prescription or dispensation of AOM (date for therapy initiation) during 2000–2014 (CPRD) or 2001–2013 (BIFAP) were included. Users of raloxifene/bazedoxifene were excluded from both databases. Five exposure cohorts were identified according to first treatment: (1) alendronate, (2) other bisphosphonates, (3) strontium ranelate, (4) denosumab, and (5) teriparatide. Participants were followed from the day after therapy initiation to the earliest of a treated VTE (cases), end of AOM treatment (defined by a refill gap of 180xa0days), switching to an alternative AOM, drop-out, death, or end of study period. Incidence rates of VTE were estimated by cohort. Adjusted hazard ratios (HR 95%CI) were estimated according to drug used.ResultsOverall, 2035/159,209 (1.28%) in CPRD and 401/83,334 (0.48%) in BIFAP had VTE. Compared to alendronate, adjusted HR of VTE were 1.05 (0.94–1.18) and 0.96 (0.78–1.18) for other bisphosphonates, and 0.90 (0.61–1.34) and 1.19 (0.82–1.74) for strontium in CPRD and BIFAP, respectively; 1.77 (0.25–12.66) for denosumab and 1.27 (0.59–2.71) for teriparatide in BIFAP.ConclusionsVTE risk during AO therapy did not differ by AOM drug use. Our data does not support an increased risk of VTE associated with strontium ranelate use in the community.

Comparative anti-fracture effectiveness of different oral anti-osteoporosis therapies based on “real-world” data: a meta-analysis of propensity-matched cohort findings from the UK Clinical Practice Research Database and the Catalan SIDIAP Database

Purpose This paper aims to compare the clinical effectiveness of oral anti-osteoporosis drugs based on the observed risk of fracture while on treatment in primary care actual practice. Materials and methods We investigated two primary care records databases covering UK National Health Service (Clinical Practice Research Datalink, CPRD) and Catalan healthcare (Information System for Research in Primary Care, SIDIAP) patients during 1995–2014 and 2006–2014, respectivey. Treatment-naive incident users of anti-osteoporosis drugs were included and followed until treatment cessation, switching, death, transfer out, or study completion. We considered hip fracture while on treatment as main outcome and major osteoporotic fractures (hip, clinical spine, wrist, and proximal humerus) as secondary outcome. Users of alendronate (reference group) were compared to those of (1) OBP, (2) strontium ranelate (SR), and (3) selective estrogen receptor modulators (SERMs), after matching on baseline characteristics using propensity scores. Multiple imputation was used to handle missing data on confounders and competing risk modelling for the calculation of relative risk according to therapy. Country-specific data were analyzed separately and meta-analyzed. Results A total of 163,950 UK and 145,236 Catalan patients were identified. Hip (sub-hazard ratio [SHR] [95% CI] 1.04 [0.77–1.40]) and major osteoporotic (SHR [95% CI] 1 [0.78–1.27]) fracture risks were similar among OBP compared to alendronate users. Both hip (SHR [95% CI] 1.26 [1.14–1.39]) and major osteoporotic (SHR [95% CI] 1.06 [1.02–1.12]) fracture risk were higher in SR compared to alendronate users. SERM users had a reduced hip (SHR [95% CI] 0.75 [0.60–0.94]) and major osteoporotic (SHR [95% CI] 0.77 [0.72–0.83]) fracture risk compared to alendronate users. Conclusion We found a 26% excess hip fracture risk among SR compared to matched alendronate users, in line with placebo-controlled RCT findings. Conversely, in a lower risk population, SERM users had a 25% reduced hip fracture risk compared to alendronate users. Head-to-head RCTs are needed to confirm these findings.

Identifying individuals with chronic pain after knee replacement: a population-cohort, cluster-analysis of Oxford knee scores in 128,145 patients from the English National Health Service.

BackgroundApproximately one in five patients undergoing knee replacement surgery experience chronic pain after their operation, which can negatively impact on their quality of life. In order to develop and evaluate interventions to improve the management of chronic post-surgical pain, we aimed to derive a cut-off point in the Oxford Knee Score pain subscale to identify patients with chronic pain following knee replacement, and to characterise these patients using self-reported outcomes.MethodsData from the English Patient-Reported Outcome Measures (PROMs) programme were used. This comprised patient-reported data from 128,145 patients who underwent primary knee replacement surgery in England between 2012 and 2015. Cluster analysis was applied to derive a cut-off point on the pain subscale of the Oxford Knee Score.ResultsA high-pain group was identified, described by a maximum of 14 points in the Oxford Knee Score pain subscale six months after surgery. The high-pain group, comprising 15% of the sample, was characterised by severe and frequent problems in all pain dimensions, particularly in pain severity, night pain and limping, as well as in all dimensions of health-related quality of life.ConclusionsPatients with Oxford Knee Score pain subscale scores of 14 or less at six months after knee replacement can be considered to be in chronic pain that is likely to negatively affect their quality of life. This derived cut-off can be used for patient selection in research settings to design and assess interventions that support patients in their management of chronic post-surgical pain.

A Bayesian Patient-Based Model for Detecting Deterioration in Vital Signs Using Manual Observations

Deterioration in patient condition is often preceded by deterioration in the patients vital signs. Track-and-Trigger systems have been adopted in many hospitals in the UK, where manual observations of the vital signs are scored according to their deviation from normal limits. If the score exceeds a threshold, the patient is reviewed. However, such scoring systems are typically heuristic. We propose an automated method for detection of deterioration using manual observations of the vital signs, based om Bayesian model averaging. The proposed method is compared with an existing technique - Parzen windows. The proposed method is shown to generate alerts for 79% of patients who went on to an emergency ICU admission and in 2% of patients who did not have an adverse event, as compared to 86% and 25% by the Parzen windows technique, reflecting that the proposed method has a 23% lower false alert rate than that of the existing technique.

Optimising classifiers for the detection of physiological deterioration in patient vital-sign data

Hospital patient outcomes can be improved by the early identification of physiological deterioration. Automatic methods of detecting patient deterioration in vital-sign data typically attempt to identify deviations from assumed normal physiological condition. This paper investigates the use of a multi-class approach, in which abnormal physiology is modelled explicitly. The success of such a method relies on the accuracy of data annotations provided by clinical experts. We propose an approach to estimate class labels provided by clinicians, and refine those labels such they may be used to optimise a multi-class classifier for identifying patient deterioration. We demonstrate the effectiveness of the proposed methods using a large data-set acquired in a 24-bed step-down unit.