Artificial Engineering (AI) and Machine Learning (ML)
ML Proofs of Concept; ML Roadmaps; Interpretable ML; Business scalability; ML Pipeline latency reduction; ML Ops Automation; IoT; Turning POCs into finished products
by leveraging a broad spectrum of evolving tools and platforms..
delivered by leaders in their field..
Jeff Henrikson – Machine Learning Principal Engineer
Jeffery Stewart – Cloud & Infrastructure Services (CIS) Executive
which has translated to impressive proven capabilities…
and delivered exceptional value to our customers…
Streaming classification to assist in clinical stroke triage
A stroke clinic certified by The Joint Commission as a comprehensive stroke center would like to maintain their good standing based on “Door to CT scan” and related metrics.
Challenges and Implications
Clinical support staff loaded the Epic EHR system with valuable information on potential stroke patients while the patients are being diagnosed in clinic.
Yet clinicians lack software support for acting on all information known about a patient at a given time.
Analyzing the EHR record for cases in which stroke triage was not as fast as desired, a small real-time data stream was extracted from the Epic EHR “globals”, including orders, ICD codes, and textual notes.
A sample corpus of the data stream was annotated by experts for cases in which text messages warning for time-sensitive circumstances was desirable.
A classifier was trained using the annotations and connected to the clinic’s existing text messaging system.
Up to the second awareness of rare but significant information in the hands of clinicians.
Audit and improvement of struggling ML effort for job descriptions
In a business providing feedback on the compensation structure of other businesses, every new customer’s job descriptions needed to be matched against a set of canonical jobs.
The matching was performed manually by a small group of expert employees
Scaling the business meant growing the job matching team
Challenges and Implications
Many years of previously onboarded jobs with titles and descriptions.
Existing machine learning efforts toward automating the job matching task had stalled for reasons that the management did not understand.
Upon investigation, we discovered that optimizing for “Precision@K” was more appropriate and as historical matches were recorded as canonical job titles, not job identifiers, a large portion of historical job matches could no longer be interpreted with certainty. We built a model for recovering the lost matching information, for use in folding together with the newly rescoped efforts of the internal machine learning team.
Our efforts resulted in growth through automation instead of hiring and dramatically increased productivity of the internal machine learning team toward business objectives.