D13

Hongyun Wu

Pin Jui Chen

Sherry Chi

Shuhei Fujimoto

​

​

PLN

City of Mesa

911 Process Improvement

• Out client - City of Mesa and the Mesa Fire and Medical Department (MFMD) runs 911 operations across the city and receives calls from both non-behavioral (physiological) and behavioral patients
• The first response team transports patients to the emergency room if they suffer any physiological condition or to appropriate behavioral facilities if they are behavioral patients without any physiological injury.
• MFMD has a behavioral health unit which specializes in assessing behavioral patients. They get dispatched to the scene upon request from the 911 dispatcher or first response team

​

• The client does not know the accurate number of behavioral patients, due to a limited amount of data on whether past patients were treated as behavioral
• Many behavioral patients without any physiological condition are currently transported to the emergency room where they wait for a few hours or even days before being sent to behavioral facilities
• This happens because
• The 911 operator fails to dispatch the behavioral unit because they might not be asking questions that help identify behavioral patients
• The first response team is unable to identify behavioral conditions and fails to request for the behavioral unit’s dispatch
• The behavioral unit is already engaged in another incident and is unable to respond to a new incident

​

The client’s goal is to enable the 911 dispatcher to identify more behavioral patients and dispatch the behavioral unit more frequently so the behavioral unit can assess and send these patients quickly to behavioral facilities.

​

​

ASU’s Objectives

• Identify the number of behavioral patients the client deals with
• Predict and discover characteristics whether 911 callers are behavioral patients
• Develop better questions the 911 dispatchers can ask to find behavioral patients

​

​

​

​

​

Datasets

• Zoom meeting with City of Mesa staffs per week
• questions or issues we have encountered during that week
• explanation and details of data set
• Expectations for the next following week
• Data cleaning , restructured and reorganized available data sets
• try to understand and dive deep into the scope of data sets by our own ways , then discuss our finding with customers to make sure we are on the right track
• Keep in touch with the client through Email
• in case we need help on data configuring or we may need more excel or csv data sets
• keeping updated for both parties

• We first marked incidents with run disposition of “Transfer Care to CPR Behavioral” or “Transport to Behavioral Health Facility” as known behavioral patients/records, and there are 1,166 of these out of ~150000 total records.
• By joining all tables together, we identified some common characteristics of known behavioral patients based on patient’s medical history, first response team’s impression on the patient, and the medication the patient had been taken.
• Going deeper with our potential characteristics of behavioral patients, we also want to know the characteristics that indicates a patient is definitely NOT behavioral patients (e.g. having physical injuries). We went back and look at the most common characteristics among all patients and especially picked out the ones that we don’t see in behavioral patients.

​

We compiled a list of indicators and contra-indicators that will later on help us build our predictive models. “+” means that an existence of this characteristics indicates the patient might be behavioral while “-” means the existence indicates non-behavioral.

This bar chart represents the counts of each medical history among known behavioral patients.

* duplications due to joining different tables together

Most patients don’t have any reported medications, but we included the next three medications (which are mostly antidepressant)

The three most common impression of behavioral patients.

We ended up only including the top 2: Behavioral/psychiatric disorder and suicidal ideation.

The most common inputs of primary complaint as well as secondary complaint for non-behavioral patients correspond with the contra-indicators that we indentified in the previous list

After EDA, we cleaned the data by removing records that have null values (might indicates no patients found, hence empty records for medical history, impression, and such) and created new binary columns for all the indicators/contraindications in the list.

We included all the identified indicators/contraindications as our predictors for the model, and duplicated the number of known behavioral records to be as many as non-behavioral records. We tried Logistic Regression and Random Forest, and we achieved the best result with Logistic Regression with a recall score of 0.78.

• Refine our predictive model based on the first response team’s feedback
• Explore additional data on census, police calls for service and opioid OD cases for further insight

​

​

​

​

​

​

• Build an ensemble model or perform clustering to aim for higher accuracy

​

​

Thank you!

Questions?