Candidate for the Lee B. Lusted Student Prize Competition
The purpose of our research is to develop patient-centered optimal resuscitation policies using electronic medical records (EMR). We seek to optimize care providers’ resuscitation decisions prior to and during acute physiological deterioration which is identified by abnormality or clinical decline in physiological measures.
Method:
We develop infinite and finite horizon semi-Markov decision processes (SMDP) for patient subpopulations to capture the uncertainty in patient health condition, patient health dynamics, the time spent in each health state and to identify optimal resuscitation actions. We conduct statistical analyses using records of adult patients admitted to the general ward of Mayo Clinic Rochester. We use vital signs to calculate a running physiological warning score to represent the health conditions. We divide the patients into subpopulations using three factors: prior medical emergency team intervention, patients’ frailty measured by the Braden skin score and patient type, i.e., medical or surgical. We estimate model parameters for each subpopulation using maximum likelihood estimates. We use the Chi-square test to test the equality of transition probabilities for the subpopulations, and the Kruskal-Wallis test to test if the sojourn times are drawn from the same distribution. The statistical tests are used to further define the subpopulations.
Result:
The statistical test results identify the key factors to classify patient subpopulations, i.e., frailty, prior medical team intervention, and patient admission type (medical or surgical). The infinite horizon SMDP models are solved by using the policy iteration algorithm. Numerical results show that the optimal policies have a control-limit structure. The optimal policies initiate resuscitation actions when the patient’s warning score exceeds a certain level which differs by subpopulation. Our finite horizon SMDP models are solved using the value iteration algorithm. Optimal policies are patient-specific and depend on the health-state and the remaining time until the end of hospitalization in the general ward. The results suggest the importance of the patient-specific resuscitation rules.
Conclusion:
Our research emphasizes the potential for improvement in bedside resuscitation that can be achieved by integrating EMR, expert opinion and analytical decision models. Our key findings are that: (i) EMR can be used to support patient-centered resuscitation decision making; (ii) key factors which identify statistically significant subpopulations are frailty, prior adverse event and admission reason, and (iii) the resuscitation rules should be patient-specific.