The long-term goal of this research is to develop a data-driven tool to guide the clinical decision of when to
repeat an inpatient laboratory test. The current proposal is to attempt the highest-risk element of this goal, which
is to develop the machine-learning approach for estimating the optimal next time to run a given test on a given
patient, taking into account the patient's clinical history, any treatments given, and current clinical status.
Clinicians often overestimate how frequently a test should be repeated, or for convenience they order tests
to be repeated every day. This wastes resources and increases the cost of care. Similarly, it can also be easy to
underestimate how often a test should be repeated, which can lead to suboptimal care, with increased morbidity
and mortality. There have been many interventions attempted over the years to reduce the frequency of repeated
tests, but they generally use subjective, expert-speci¿ed rules that set minimum testing frequencies for certain
clinical scenarios. These attempts are laudable, but they are rather blunt instruments for the problem, because
they cannot adapt to the speci¿c and varying needs of a patient's pathophysiologic state and treatment regimen,
and they don't affect tests that aren't being run frequently enough.
The key to providing data-driven timing guidance is the ability to estimate from data the rate at which a speci¿c
observed result will go stale, meaning that it is old enough that the estimated current value is too uncertain to
be used for decision making. The mathematically optimal time to order a test repeat is exactly when the most
recent value reaches that level of undesirable uncertainty. In hospitalized patients, the rate at which the latest
value of a test goes stale changes over time due to many factors, and therefore, so does the optimal time for the
next test. And while it is known how to tell in hindsight what the optimal repeat time would have been for a test,
it is unknown how accurately it can be predicted for the next, future sample.
This project assesses the technical feasibility of providing this predictive guidance at institutional scale, with
the following speci¿c aims:
Aim 1: (Accuracy) Develop and assess the computational accuracy of personalized timing guidance
under moderate-scale data conditions. We will develop models to provide timing guidance for the 125 most-
repeated inpatient numeric lab test, and assess their accuracy under moderate-scale data conditions.
Aim 2: (Scalability and Utility) Determine the computational scalability and need for personalized tim-
ing guidance. We will develop and assess the speed vs. approximation trade off needed to scale up the timing
models to full institutional data size. We will then assess historical orders to quantify the gain to be achieved by
using personalized timing guidance under our best model.