Healthcare Associated Infections

• Healthcare associated infections (HAIs) associated with increased morbidity and mortality
• Five of the most common HAIs are:
  • Clostridiodes difficile – causes severe damage to the colon, can be fatal
  • Bloodstream infection (sepsis) – estimated mortality rate of 15-30%
  • Urinary track infection – low mortality but associated with multi-drug resistance and significantly longer hospital stays
  • Healthcare acquired pneumonia – mortality rate of 40-70%, increasing dramatically with age
  • Surgical site infection – significantly increases length of stay

Estimation methodology

Step 1: Hospital prevalence

Hospital prevalence, P, estimated as

P = r \times Beta(n_{\text{obs}}, N-n_{\text{obs}}+1) + (1-r) \times Beta(n_{\text{obs}}+1, N-n_{\text{obs}})

Estimation methodology

Step 2: Estimation of hospital incidence

Hospital incidence, I, calculated as

I = P \frac{\text{LA}}{\text{LOI}}

where

• P is the hospital prevalence from step 1,
• \text{LA} is the mean length of stay and
• \text{LOI} is the length of infection.

Estimation methodology

Step 2a: Estimation of length of infection

No data on length of infection, only LOI_\text{pps}, the length of stay until the date of survey.

We can calculate

P(LOI_\text{pps} = 1),

the probability a patient is in the first day of their HAI. Then, approximate LOI with

E[LOI] = 1 / P(LOI_\text{pps} = 1).

For small sample sizes, results has shown this is biased, so we use a mixture of this estimator and the empirical mean.

Estimation methodology

Step 3: Estimation of population incidence

Calculate population incidence as

I_{\text{pop}} = I \times N_{\text{discharges}}.

For us, N_{\text{discharges}} = 3,713.513, which is 60% of the total admissions in the year.

Estimation methodology

Step 4: Stratification by age and sex

Use a multinomial likelihood with a Dirichlet prior, with weights taken from the number of cases in each age/sex category.

A psuedocount is added to each strata (0.001\sum \text{weights}) to ensure likelihood can be calculated with empty strata

Key results

Carbapenemase-producing Enterobacterales

• Gram-negative bacilli occurring naturally in the GI tract
• Resistant to carbapenem antibiotics
• Several different carbapenemases genes in CPE:
  • Imipenemase (IMP)
  • Klebsiella penumoniae carbapenemase (KPC)
  • New-Delhi metallo-β-lactamase (NDM)
  • Verona integron-encoded metallo-β-lactamase (VIM) and
  • Oxacillinases (OXA)
• For Enterobacterales, IMP(-4) most commonly found in Victoria (28% of all cases)
• Assumed to be colonised indefinitely

Carbapenemase-producing Enterobacterales

• Surveillance officially established in Victoria in 2015, became notifiable in 2018
• Since becoming notifiable, notifications have been steadily increasing

Health Services in Victoria, Australia

• Health services are individually governed
• Centralised guidelines for reporitng
• Have their own IP results, testing/screening requirements, admission & transfer patterns
• Are often spread across multiple campuses which operate independently but are administered centrally

Data sources

Victorian Admitted Episodes Dataset (VAED):

• Line list of every hospital admission (public and private)
• Reports demographics, conditions, treatments, timing, location

Public Health Event Surveillance System (PHESS):

• Records information on notifiable disease events
• Includes patient demographics, date of notification, date of test
• Operational system, used for contact tracing and outbreak analytics

Control Population

To fairly compare between CPE+ and CPE- patients, construct a series of comparator populations which increasingly closely approximate the CPE+ cohort

1. Random subset: patients are randomly sampled from the population
2. Campus and time: Patients are matched to CPE-positive patients based on the hospital campus and the quarter-year (i.e. 3-month period) of admission
3. Campus, time and age: As per Cohort 2, plus the inclusion of five-year age band
4. Campus, time and comorbidities: As per Cohort 2, plus the inclusion of age-adjusted Charlson Comorbidity Index category.

Survival Analysis

Use the Kaplan-Meier estimator:

\hat{S}(t) = \prod_{i:t_i \leq t} 1 - \frac{d_i}{n_i}

where

• d_i is the number of events that have happened up to time t_i, and
• n_i is the number of individuals known to have survived up to time t_i.

Statistical methodology

To calcualte the time until next admission, we set the time 0 to be the time of discharge from a health service where CPE was diagnosed

Statistical methodology

Time of event is defined to be the time at which a patient was readmitted to a new health service Time of censoring is either death (if known) or the end of the data (30 November 2019), meaning patients have unequal study lengths.

Other ongoing work

• Machine learning methods to identify risk factors for hospital infections
  Leong Zhuan Kee, Vis; Ewilly Liew
  
• Estimating rates of contact in the absence of local data, conmat
  Nick Tierney; Nick Golding
  
• Understanding how non-linear dimension reduction warps your data
  Jayani Lakshika; Di Cook
• Predicting extreme weather events using short- and long-term climate drivers
  Kate Saunders; Jarryd Chapman
• Approximating patient movement patterns with piecewise constant networks
  David Wu; Andrew Stewardson
  
• Developing non-network based network layout algorithms
  Krisanat Anukarnsakulchularp; Di Cook

…and more.