population health - know your audience€¦ · population health - know your audience katja...

Population Health -

Know your Audience

Katja Behrendt – Innovation Delivery Lead

Slides from NHS Expo 2018

1

Suzuki SV650 motorbike wiring diagram

2

http://www.longwoods.com/content/20144

3

http://www.goinvo.com/features/determinants-of-health/

4

Analytic tools

5

JSNA

The Joint Strategic Needs Assessment (JSNA) is a process through which the council's social care services (Education, Social Care and Wellbeing) work together with public health and NHS services to assess the needs of a population and determine priorities for commissioning services

The first step in the cycle is to undertake a Joint Strategic Needs Assessment (JSNA) for the defined population of the ACS. Each local authority within the geographical boundaries of the ACS will already have produced a JSNA for its population, so these documents would be a good place to start. The JSNA for the ACS should be undertaken jointly by NHS staff and local authorities and their public health teams, using a combination of NHS and local authority data, where possible linked at the person level. It will determine the health and care needs of the population and important segments within it. These segments can be demographic, geographic, and segments based on condition categories, the distribution of long term conditions or other disease and multi-morbidity groupings. The involves understanding the wider determinants of health that are acting upon the population; analysing health inequalities; benchmarking

6

against peers; measuring 'wellness' and its distribution; undertaking geospatial analysis; and using a range of simulations, system dynamics and other techniques to predict a range of demographic changes in the population and test their impact. Visualisations can help bring the data to life, including atlases of variation, patient timelines (“Theographs”) and patient pathways (“Bow Tie” diagrams)

6

Very High 0-0.5%

High 0.5 – 5%

Moderate 5 – 20%

Low 20 – 100%

Exp

erie

nce

of

X o

f ti

me

Y

Size of population

Risk Stratification

Data for risk stratification

• clinical/diagnostic• prior utilisation/cost• socio-demographic• pharmacy• health status/functionality• social care• social e.g. social isolation, living in crowded housing

7

• Clinical perspective: • using clinical knowledge of patients

• Threshold modelling: • picking patients according to a rule e.g. >65 with 2+ hospital

admissions in previous 12 months• Predictive modelling:

• including multiple regression, decision trees, AI

Clinical perspective: using clinical knowledge of patientsAdvantages: acceptable/preferred for clinicians, can identify patients who are impactableDisadvantages: no more accurate than chance, not able to scan populations repeatedly, cannot predict for patients not seen, bias

Threshold modelling: picking patients according to a rule e.g. >65 with 2+ hospital admissions in previous 12 months

Advantages: data easily accessible, low costDisadvantages: less accurate than chance due to regression to mean, not accurate at predicting future risk

Predictive modelling: including multiple regression, decision trees, AIAdvantages: most accurate (twice as accurate as threshold models)Disadvantages: relatively expensive, requires analytical support e.g. Commissioning Support Unit, commercial

8

Predictive models are recommended as they are the most accurate

Examples of predictive modelling tools available are:• Patients at Risk of Re-hospitalisation: commissioned by

Department of Health• Combined Predictive Model: developed by King’s Fund • Adjusted Clinical Groups: developed by John Hopkins University• QAdmissions: developed by ClinRisk LTD,

These models are reasonably similar in terms of their predictive performance

PARR - https://www.kingsfund.org.uk/projects/predicting-and-reducing-re-admission-hospitalcommissioned by Department of Health, developed by the King’s Fund, free to use, can only be used with patients admitted to hospital in last 24 months.

The combined model -https://www.kingsfund.org.uk/sites/default/files/field/field_document/PARR-combined-predictive-model-final-report-dec06.pdfdeveloped by King’s Fund, free to use, can be used for patients with no admissions in last 24 months

ACG - https://www.sollis.co.uk/press-release/acg-system-calibrated-for-uk/classify each patient into one of 93 ACG categories, uses inpatient and outpatient diagnoses

Q Admissions - http://www.qadmissions.org/ run off routinely collected data already in GP computer systems and validated

9

in primary care setting

Model Training/Validation cohort sizes (number of individuals) -Validation set Performance (c-statistic)Brave AI 414,372/138,123 - 0.97Gao 1,493,526/1,493,526 - 0.83Sussex Key Events Predictor tool 658,000/122,000 - 0.82Wang 2,700,000/1,800,000 - 0.81Combined Predictive Model 280,000/280,000 - 0.8ACG 4,630,000/4,700,000 - 0.8Baker Not recorded/Not recorded - 0.79PEONY 90,522/90,879 - 0.79Optum HealthNumerics-RISC 414,372/138,123 - 0.79Billings 1,836,099/Not recorded - 0.78Devon Predictive Model 577,906/144,477 - 0.78Qadmissions 2,849,381/3,700,000 - 0.77MoSaiCO 146,949/147,654 - 0.77Louis 200,000/50,000 - 0.77Lopez-Aguila 28,430/Not recorded - 0.76Damush 1,041/1,000 - 0.74Inouye ,932/1,987 - 0.73Crane 12,650/450 - 0.7EARLI 3,302/500 - 0.69Boult 2,942/2,827 - 0.69Reuben 5,138/5,138 - 0.68CARS 411/1,054 - 0.67Mazzaglia 2,470/2,926 - 0.67Freedman 1,873/1,872 - 0.63Sherbrook 2,307/532 - 0.6SPARRA-version 3 3,506,796/Not recorded - 0.6DxCG Not Recorded/Not recorded Not RecordedRoos 1518/1518 Not RecordedWPM 298077/74114 Not Recorded

9

Mostly Healthy Adults

MostlyHealthy

+65

MostlyHealthy

+75 (Frail)

LTC(adult)

LTC(>65)

2 LTC(adult)

2 LTC(>65)

Cancer

Acute 10411 1219 857 5004 1606 5015 8277 7530

CHS 461 278 162 521 196 453 954 255

SC 530 61 122 850 554 1252 3649 1403

MH 1096 39 22 845 42 532 226 253

GP 3331 262 120 1706 374 1454 1503 754

RX 682 103 80 1139 397 1649 2382 826

Pop(‘000)

27,834 1,322 564 6,060 1,590 2,464 2,555 1,244

Spend(£m)

16,724 1,987 1,382 10,794 4,882 8,538 15,517 11,179

Segmentation

10

William Francis "Willie" Sutton, Jr. (June 30, 1901 – November 2, 1980) was an American bank robber. During his forty-year robber career he stole an estimated $2 million, and he eventually spent more than half of his adult life in prison and escaped three times. For his talent at executing robberies in disguises, he gained two nicknames, "Willie the Actor" and "Slick Willie". Sutton is also known as the namesake of Sutton's law, although he denied originating it.

Sutton's law states that when diagnosing, one should first consider the obvious. It suggests that one should first conduct those tests which could confirm (or rule out) the most likely diagnosis. It is taught in medical schools to suggest to medical students that they might best order tests in that sequence which is most likely to result in a quick diagnosis, hence treatment, while minimizing unnecessary costs. It is also applied in pharmacology, when choosing a drug to treat a specific disease you want the drug to reach the disease. It is applicable to any process of diagnosis, e.g. debugging computer programs. Computer-aided diagnosis provides a statistical and quantitative

11

approach.A more thorough analysis will consider the false positive rate of the test and the possibility that a less likely diagnosis might have more serious consequences. A competing principle is the idea of performing simple tests before more complex and expensive tests, moving from bedside tests to blood results and simple imaging such as ultrasound and then more complex such as MRI then specialty imaging. The law can also be applied in prioritizing tests when resources are limited, so a test for a treatable condition should be performed before an equally probable but less treatable condition.The law is named after the bank robber Willie Sutton, who reputedly replied to a reporter's inquiry as to why he robbed banks by saying "because that's where the money is." In Sutton's 1976 book Where the Money Was, Sutton denies having said this.[1][2]1. A similar idea is contained in the physician's adage, "When you hear

hoofbeats, think horses, not zebras."

11

Impactibility modelling

• Predicts which high-risk patients are most likely to be responsive to the preventive care intervention being offered.

• Approaches to impactibiliy modelling include:• Ambulatory Care–Sensitive Conditions• Gap analysis• Excluding patients unlikely to respond*

*In the States, some providers systematically exclude patients if English is not their first language, or if they have drug & alcohol problems or mental health problems. This approach would be considered unethical and probably illegal in the UK

12

Ambulatory Care–Sensitive Conditions

Increase the impact of predictive risk models by giving priority to patients with certain diagnoses amenable to upstream care

heart failure and epilepsy

13

• Defined as an evidence-based intervention that would be expected for this individual patient but has not been delivered • such as a test, immunisation or treatment

• Prioritising patients with a high gap-score• Prioritising patients with a high weighted

gap score

Gap analysis

• Prioritising patients with a high weighted gap score • same as above but

weighted different gaps depending on how important they are to a

14

patient’s care

14

Variation

15

Variation

16

Impactable populations

17

Pathways

18

Effective (personal) value • is each patient in the health economy receiving

interventions that addresses their personal preferences? • preference-sensitive conditions (e.g. back

pain, prostate cancer) • are patients offered decision support to help

them make a joint decision with their clinician?

21

Efficient (technical) value

• are the services delivered as efficiently as

possible?

• is there a way of improving the outcomes it

achieves for the same or lower cost?

22

Economic (allocative) value

• is the health economy investing in the right mix of

services to address the needs of its population?

• between services

• within services

23

Evaluation

27

0.0

0.1

0.2

0.3

-12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 1 2 3 4 5 6 7 8 9 10 11 12

Nu

mb

er

of

eme

rgen

cy h

osp

ital

ad

mis

sio

ns

pe

r h

ead

pe

r m

on

th

Month

Intervention

Start of intervention

Propensity score matched cohort studies

28

0.0

0.1

0.2

0.3

-12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 1 2 3 4 5 6 7 8 9 10 11 12

Nu

mb

er

of

eme

rgen

cy h

osp

ital

ad

mis

sio

ns

pe

r h

ead

pe

r m

on

th

Month

Intervention



29

0.0

0.1

0.2

0.3

-12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 1 2 3 4 5 6 7 8 9 10 11 12

Nu

mb

er

of

eme

rgen

cy h

osp

ital

ad

mis

sio

ns

pe

r h

ead

pe

r m

on

th

Month

Intervention



30

0.0

0.1

0.2

0.3

-12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 1 2 3 4 5 6 7 8 9 10 11 12

Nu

mb

er

of

eme

rgen

cy h

osp

ital

ad

mis

sio

ns

pe

r h

ead

pe

r m

on

th

Month

Control Intervention



31

•PDSA cycles

•Learning health systems

•Evaluations built in to programmes• Track and evaluate•Measure impactibility as we go

32

Full evaluations EconomicOutcomeMixed methods

33

population health - know your audience€¦ · population health - know your audience katja...

Documents