Patient Journey Record(pajr) - Jing Su
- 1. Patient Journey Record (PaJR) Online Prediction System Jing Su, Lucy Hederman, Atieh Zarabzadeh, Dee Grady, Carmel Martin, Kevin Smith, Carl Vogel, Enda Madden, Brendan Madden Trinity College Dublin, UCC, PHC Research, GroupNos HISI 2011 1
- 2. Avoidable hospitalisations • PaJR is a telephone service that targets avoidable hospitalisations • Most hospital admissions • are in older, sicker people with multiple diseases and conditions • are unpredictable in the short term with current systems HISI 2011 2
- 3. $/day $5,000 $1,000 $100 Early Alerts of deterioration helps prevent the patient $10 entering expensive treatment $1 Self care Complicated Complex Hospital HISI 2011 3
- 4. The PaJR Service Michael Anon HISI 2011 4
- 5. The PaJR Survey Michael Anon HISI 2011 5
- 6. The PaJR approach • Use lay callers • Ask the questions that predict hospitalisation • Not disease specific • Intervene early • Alert accurately … • ~ 50% reduction in hospital admissions in pilot sites HISI 2011 6
- 7. PaJR Prediction Service Michael Anon • Predicts patient deterioration based on record of call. – Self-rated health, taking meds,… – Brief text entries • Uses a predictive model learned from examples of calls leading to unplanned events. HISI 2011 7
- 8. Machine Learning ML: automatically induce from the examples a model that accurately predicts new cases. HISI 2011 8
- 9. Simple predictive model: a decision tree A - decision node A yes no - leaf node K=y (K means B B UnplannedEvent) <n ≥n <p ≥p C A sample decision tree K=y K=n K=y on UnplannedEvent yes no Eg: A = “her sister”; B = AvgWordLength; K=n K=y C = takingMeds HISI 2011 9
- 10. Machine learning • ML Tools such as Weka or Timbl provide algorithms to produce prediction models from examples (“training data”). • The examples must be presented to ML tool as a collection of features. • Expertise and skill is needed to identify / derive / represent features of examples that might predict the outcome. HISI 2011 10
- 11. PaJR‟s Current ML • Predicts unplanned events, urgent unplanned events, self rated health. • Uses decision trees. • Weights false negatives 500 more costly than false positives – A missed deterioration is bad. – An inappropraite alert to a carer to call a patient is OK. HISI 2011 11
- 12. Accuracy • Predicting urgent unplanned events True False (UUE) Negative 1091 4 • Training data (OK) – 1621 phone calls Positive 23 453 – 27 urgent unplanned (UUE) events • False Negatives cost 500 times FPs • Fewer than 1/3 of the calls are incorrectly prioritised • Under 1/6 of the calls that should be prioritised are not. HISI 2011 12
- 13. Error Analysis • False positive rate is worth further investigation: – ML predicts an urgent unplanned event. – No urgent unplanned event occurred. – But is that because the PaJR caller intervened (with advice, referral, comfort, …) and averted the event ML predicted? • Analysing FP cases, we found evidence of some intervention in a small number of cases. – Further work needed. • More significantly, UUEs were rarely (6/27) „anticipated‟ by lay callers (they didn‟t intervene), whereas ML predicted 23 of them. HISI 2011 13
- 14. Challenges • Data – ML requires lots of examples of each outcome. Thanks to PaJR the number of unplanned events among the users is declining. • Features – We have lots of data for each case but it takes time and skill to identify features predictive of deterioration. • Prediction Engine Pipeline – The management of multiple cases, multiple models, etc. HISI 2011 14
- 15. Benefits of Machine Learning • Compared with static rule-based alerts – ML allows identification of features that emerge as predictive of deterioration. – ML uses evidence from data on real patients. – ML can be easily transferred to new settings and new services – ML adapts over time • Compared with experienced callers without ML – ML allows high accuracy, high volume at low cost. – ML will identify features across callers, across time, etc. – ML has perfect memory – callers go on leave, move on. HISI 2011 15
- 16. Thank you Any questions? Lucy Hederman hederman@tcd.ie HISI 2011 16
- 17. HIDE???? Machine Learning Pipeline Querier Language ML Parser Algorithm Existing surveys Training data Hours Parsed Decision Database CSV Features Tree New survey Test data Predicting Unplanned Moments Events / SRH (< 1 minute) Off-line training and online predicting! Update Decision Tree model at regular intervals Qualitative feedback HISI 2011 17
Editor's Notes
- #3: Currently estimates for the bill for avoidable re-admissions in the US is $15 billion with a corresponding figure of £1.5 billion published by the UK Department of Health in 2010. From 2013 in the US, hospitals will be financially penalised for every re-admission of a Medicare patient within 30 days of their discharge. The number of patients over 65 years of age admitted to hospital in the US is 13 million per year. Currently, 20% of these patients are re-admitted within 30 days. Initial trials of the PaJR RAP Service have shown a 50% decrease in re-admissions. This would make the RAP Service very cost effective for US hospitals to deploy. Reducing avoidable hospitalisationsOne night in an Irish hospital costs €1000As a locum GP I’ve visited elderly patients with multiple chronic illnesses – I’ve had to admit patients as the degree of their deterioration is unclearIn 2010 the worldwide market (US, EMEA and APAC) for home health monitoring of “welfare diseases” such as diabetes, cardiac arrhythmia, sleep apnea, asthma and chronic obstructive pulmonary disease (COPD) was worth about €7.6 billion ($10 billion). The market for remote monitoring services is set to grow rapidly through 2014 based on current reports. Estimates for the cost savings achievable through remote monitoring of patients in the US by 2014 range from $2 billion and $6 billion.
- #7: The PaJR Service identifies patients at immediate risk of hospitalisation and alerts their GPsTimely interventions avert deteriorations.PaJR reduces admission rates by 50%.
- #16: Emergent feature – diarrhoea, falls; “cold”; Some of these may feed back to service and discharge planning.