Last revised: March 31st, 2021

Revision History:

Sam Cymbaluk - 400138646

Shawn Esseltine - 400145888

Daniel Rubinstein - 400111892

Nikhail Singh - 400069771

Kristie Tom - 400018028

• Modules and Module Interfaces:
  • UI Modules
    • Prediction Module
    • Admin Module
  • Backend Modules
    • Rest API Spec Version v1
    • Database Module
    • Triage Controller
    • ML Training Module
    • Simulation Module
• Dependency Diagram
• Likely Changes
  • Data
  • Behaviour
    • UI
    • API
    • Triage Controller / Simulation Module
    • ML Module
  • Hardware Specifics
  • Security Requirements
• Unlikely Changes
  • Data
  • Behaviour
    • UI
    • API
    • Triage Controller / Simulation Module
    • ML Module
  • Hardware Specifics
  • Security Requirements
    • UI Authentication
    • API Authentication
    • Data Usage Security
    • Data Storage Security
• Problems Related to Problem Domain
• List of Changes to SRS
• Removed Requirements

A page that creates predictions for the optimal safe schedule given the clinic constraints.

• REST API
• Angular

The UI interface for the prediction module is defined graphically as a form that when inputted the following constraint parameters:

• Prediction range: int
• confidence level
• interval date ranges (intervals within prediction can be of different length)
• Target service % for patient α_i: int, for each triage category denoted as α_i
• target time window for patient α_i:int for each triage category denoted as α_i
• (optionally) a file with the initial patient backlog

It returns the minimum number of required slots for all triage categories (i.e. urgent patient, semi-urgent patient, etc.) needed to see the expected number of patients in each category within their category's respective timeframe; the confidence interval around that number; and a breakdown of the expected slots per class;

Below are the major UI methods used to get the predictions:

From the HTTP response, store the data in a "global" service to allow inter-component communication.

Returns

• void

Makes an http request to the REST API to get the prediction given the url.

Parameters

• url <string> - the request url

• body <string> - the POST request body

Returns

• Object - the http request response body

• The user input updates the state of the angular component (which is used to create the prediction module). Once all the fields are validated, for their correct type
• The user inputs must be type validated using REGEX matching
• Dynamic reloading is done via angular observables

Allows any technical or administrative staff to upload new data, and change any default behaviours.

• REST API
• Angular
• angular-file-uploader

The UI interface for the admin module is defined graphically in 2 parts:

i) Data uploading

More up-to-date data can be uploaded to keep the predictions relevant. Therefore the following can be uploaded in a dropbox:

• Most recent patient waitlist
• Past patient visitation patterns

ii) Configuration settings

Changing any default values can be done via a form that intake the following parameters:

• The default prediction constraint values (mentioned above)

The file stored in the module state as a private variable is sent in the post body to the REST API

Returns

• void

• Dynamic reloading is done via angular observables

Some description

• Triage Controller
• Flask
• api.resources.predict
• json
• math
• model.load_and_predict
• scipy

The flask API interface can be defined via a REST schema

Given the following prediction constraint parameters, fetch the current wait list, get the patient arrival distribution, and then calculate the number of slots to allocate per triage class that maximizes the number of patients seen while satisfying the prediction/query parameters constraints. See the Secrets (Implementation Details) section for details.

Below is some JS pseudocode for type defining

type Number_By_Triage_Class = {
	urgent: number
	'semi-urgent': number
	standard: number
}

type Prediction: = {
	'url': string,
        'predictions': [triage-class: {
                         'start': string,
                         'end': string,
                         'slots': number
        }]
}

• When the endpoint is invoked, load_and_predict_patient_total is invoked as a subroutine and returns the total number of patients predicted. From this, scipy linear optimization is used to find the minimum number of slots needed so that patients can be seen within the appointment deadlines. If no solution is found, a greedy algorithm for the number of slots is implemented, favouring patients by triage class priority. The slot distribution result is returned.

Send the past appointment history by triage class data to base predictions off the most recent data history. Restricted to privileged users.

• Contents of the file are validated

Get a list of all the prediction models for a particular clinic. Restricted to privileged users.

Update a clinic’s prediction simulator to use a particular model. Restricted to privileged users.

• Prediction models have a creation date. Using this date, a cron job removes models that exceed a configurable retention time. This way, the list of models does not grow infinitely.

Display triage classes for a given clinic.

• DataBase module used to access data.

Update a triage class for a given clinic.

• DataBase module used to access data.
• Triage class changes are validated.

A Postgres database that stored all the persistent data for Triage

• Postgres

The database interface consists of the tables and views.

• There is an index on User.username for faster user lookups
• There is an index on Schedules.clinic_id to more efficiently find cached schedules
• There is an index on Models.clinic_id to more efficiently get the model for a clinic
• There is an index on TriageClasses.clinic_id to more efficiently get clinic triage classes

The controller that provides the functionality exposed by the REST API

• Simulation Module
• ML Training
• Database

Construct a new Triage Controller

Returns

• A TriageController object

Construct a Simulation

Parameters

• clinic_id <int> - The clinic to run the prediction on
• parameters <Dict> - The prediction parameters (all the triage class durations and proportions, and the number of times to run the simulation)

Returns

• A Schedule object that contains the required number of slots for each triage class

Get a clinic’s DataManager

Parameters

• clinic_id <int> - The clinic to get the DataManager of

Returns

• A DataManager object with the following functions
  • getWaitlist
  • updateWaitlist
  • getPastAppointments(start, end)
  • addAppointments(appointments)

Get a clinic’s ModelManager

Parameters

• clinic_id <int> - The clinic to get the ModelManager of

Returns

• A ModelManager object with the following functions
  • listModels
  • getModel
  • getModel(id)
  • trainNewModel(hyperparams)

• The controller schedules recurring training sessions on the updated data for each clinic
• The predict function uses the database to cache prediction results

A transient module that performs a training session

• Database

This module initialized with the following details

• Clinic ID
• Training hyperparameters

It then performs the training session and writes the model to the database module.

• Uses a GPU for training
• Preprocesses the training data
• Calls database to get training data

Performs simulation using predicted data to produce a schedule

• Controller

sim/README.md documentation of the interface

• Uses a ML model loaded from the database to predict future patient volumes
• Design decision doc

Send the most current clinic waitlist to base predictions off of. Restricted to privileged users.

Originally determined that the client could upload a waitlist but the client has chosen not to give us a waitlist and to base predictions off of only seen patients.

• Contents of the file are validated