MathWorks’ Johanna Pingel talks with Senior Editor Bill Wong about how engineers can use artificial intelligence to optimize their workflows.

Engineers are increasingly seeking to integrate AI into their projects, both to improve their results and remain ahead of their profession’s digital curve. To successfully integrate AI, engineers should make sure they understand what, exactly, AI is in the first place, and how it can fit into their current workflow. It may not be as straightforward as they first believe.

In this Q&A, Electronic Design’s Senior Editor Bill Wong talks with Johanna Pingel, product marketing manager at MathWorks, about how engineers can integrate AI into their projects, and how it can ultimately be used to optimize a complete workflow.

How do you define AI from an engineering perspective?

When engineers discuss AI, they’re usually focusing on AI models, but AI is much more. It’s an often-nebulous term that describes an operational strategy supported by machine learning. In engineering terms, the concept of “AI” actually spans four steps within a workflow: data preparation, modeling, simulation and testing, and deployment.

Which step or steps are most important when incorporating AI into a workflow?

Each step is important. It’s crucial for engineers to remember that, because they often expect to spend most of their time on the second step—developing and fine-tuning AI models.

While modeling is undeniably a key part of the process, it’s neither the beginning nor end of the integration journey. If any step in practical AI implementation is most important, it’s the first, which is data preparation. It’s critical to uncovering issues early on and knowing which parts of the workflow to focus on to achieve the best results.

Of course, the most important step will depend on the specific application. But when in doubt, start with the data.

What else should engineers consider before incorporating AI into their workflow?

Engineers should recognize the value of their existing knowledge. When developing an AI workflow, many believe they lack the skills necessary to incorporate AI into their projects, and that’s rarely true. They have inherent knowledge of the problem they’re trying to solve, and access to data preparation and modeling tools that can help them leverage that expertise, even if they’re not AI experts.

They should also keep in mind that AI is only one part in a much larger system, and all parts must work together for its implementation to be successful.

Walk us through the four steps to developing a complete AI-driven workflow. What role does each step play in successfully incorporating AI into a project?

As mentioned, the first step, data preparation, is arguably the most important. Often, when deep-learning models don’t work the way they’re expected to, engineers focus on the second stage—fine-tuning their model, tweaking its parameters, and implementing multiple training iterations. They fail to realize that to be effective, AI models need to be trained on robust, accurate data. If an engineer gives the model anything less, they will get no insight from their results, and likely spend hours trying to learn why the model isn’t working.

Instead, engineers are better served by focusing on the data they’re feeding into the model. Preprocessing the data and ensuring it’s correctly labeled helps ensure the model will be capable of understanding the data.

For example, engineers at construction equipment manufacturer Caterpillar have access to high volumes of field data generated by their machinery’s industry-wide use, but recognize that the sheer volume of data can interfere with their model’s effectiveness. To streamline the process, Caterpillar uses MATLAB to automatically label and integrate data into their machine-learning models, resulting in more promising insights from their field machinery. The process is scalable and gives Caterpillar’s engineers the freedom to apply their domain expertise to the company’s AI models without forcing them to become AI experts themselves.

Once the data is prepared, how important is the next step of the workflow—namely, modeling?

Assuming the data-preparation stage is complete, the engineer’s goal at the modeling stage is to create an accurate, robust model capable of making intelligent decisions based on the data. This also is the stage where engineers should decide what form it should take, whether that’s machine learning like a support vector machine (SVM) or decision trees, deep learning like neural networks, or a combination of the two; choosing whichever option produces the best result for their application and business needs.

It’s important for engineers to have direct access to multiple workflow algorithms, such as classification, prediction, and regression. In addition to providing more options, this allows them to test their ideas with prebuilt models developed by the broader community, and potentially use one as a starting point.

It’s also crucial for engineers to remember that AI modeling is an iterative step within the workflow. They must track whatever changes they’re making throughout, as it can help them identify the parameters that increase the model’s accuracy and create reproducible results.

Now that we’ve prepared our data and set up a model, where does simulation and testing come in?

This step is the key to validating that an AI model is working properly and, more importantly, working effectively with other systems before it’s deployed in the real world. Engineers must keep in mind that AI models are part of a larger system and must work with all other pieces of that system. Consider an automated driving model: Not only must the engineers design a perception system for object detection like stop signs, other cars, and pedestrians, but it must be integrated with other systems like controls, path planning, and localization to be effective.

The testing stage is essentially an opportunity for engineers to ensure the model they developed is accurate, and the best way to test that model is through simulation, using virtual tools such as Simulink.

At this stage, engineers should ask themselves questions to ensure their model will respond the way it’s supposed to, regardless of the situation. What is the model’s overall accuracy? Is the model performing as expected in every scenario? Does the model cover all edge cases?

By testing for accuracy via simulation, engineers can verify their model’s reliability under all anticipated use cases, avoiding costly redesigns that drain both money and time once a model is deployed.

So, we’re finally ready to deploy our model. What role does AI play in this final step?

The deployment stage is no longer about the model, which has now been verified to be processing and extracting accurate insights from prepared data, but about the hardware it’s being applied to and the language used. For example, a model can run directly on a GPU, and automatic generation of highly optimized CUDA code can eliminate coding errors often introduced to the GPU through manual translation.

Engineers should keep this stage in mind throughout the process, ensuring they ultimately share an implementation-ready model compatible with their project’s designated hardware environment, which can range from the cloud, to desktop servers, to FPGAs.

Here, too, the right tools can make this stage easier. Flexible software capable of generating the final code in all scenarios enables engineers to deploy a model across multiple environments without forcing a rewrite of original code.

The post How to Integrate Artificial Intelligence into Your Workflow appeared first on Artificial Intelligence.

Big data analytics is no longer a nice thing to have for enterprises: It’s now mission-critical.

In 2019, Veritas said, “In just a few years, big data has advanced from scattered experimental projects to achieve mission-critical status in digital enterprises, and its importance is increasing. According to IDC, by 2020, organizations able to analyze all relevant data and deliver actionable information will earn $430 billion more than their less analytically oriented peers. Big-data analytics, once performed on an occasional basis, are now performed daily at many enterprises, including, Amazon, Walmart, and UPS.” 

Yet organizations continue to experience difficulty in trying to operationalize it. 

Gartner defines big data operationalization as, “the application and maintenance of predictive and prescriptive models. Both clients and vendors are placing an emphasis on the importance of moving data science out of a prototype environment and into a state of production and continuous improvement.” 

In other words, to operationalize big data, you have to move it out of the test sandbox and into an active role in the business.

The most active roles for big data in the business to date have been in decision support. 

• Consumer buying patterns from web-based data inform retailers about which products are moving fastest, who is buying them, and where they are being bought.
• Diagnostic analytics systems enhanced by machine learning inform medical practitioners about the most likely diagnoses and treatments for certain conditions.
• Sensors placed along tram tracks and on key pieces of equipment inform cities which areas in their physical tram systems require immediate or near-term repair so the system will not fail.

All of these examples illustrate a first tier of big data analytics deployment in that they use unstructured big data and their role is in providing static reports to managers that can be acted upon.

Using analytics in daily workflow

However, when you fully operationalize analytics, there is also a second-tier active stage of engagement in which companies embed big data analytics directly into the daily workflows of their operations. In these instances, the analytics continue to inform decisions but they also automate certain tasks in company workflows based upon the intelligence they glean from data.

A great example of system automation in operations is decision-making in bank lending. For many years, software programs assessed a loan applicant’s credit worthiness and determined a “lend” or “don’t lend” decision and a loan rate that took into account the loan applicant’s credit status, the size of the loan, and the amount of risk. 

The lending supervisor still has the final say, but in essence the lending software has made the decision.

We can extend this model into the area of maintaining a city tram system.

Internet of Things (IoT) Sensors are attached to key pieces of track and equipment. The sensors can detect signs of failure in these physical components before failure occurs. Data is collected, and reports are generated for supervisors, who then organize preventive maintenance tasks and routes.

Now, what if these analytics could be operationalized even further? For instance, an analytics system picks up big data in real time from IoT sensors dispersed throughout the city’s transit system. The system analyzes this data and produces maintenance reports for supervisors—but it also interfaces with a work order planning system that organizes maintenance work by location and sequences work orders for crews.

These work orders could be dispatched directly to maintenance crews or the organization could choose to have a human supervisor review and then authorize the work orders before issuance.

By integrating big data analytics into day-to-day task loads that go beyond just reporting (i.e. tier-two operationalization), organizations can achieve greater returns from their analytics and big data investments.

This is more important than ever because just last year Venturebeat reported that 87% of data science projects still never make it into production.

Going forward, we can’t afford this level of failure for big data and analytics. Operationalizing it in business workflows as well as in static reports is all the more vital. 

The post Why it’s important to operationalize big data into daily tasks appeared first on Artificial Intelligence.

Bon Secours Charity Hospital, a three-hospital health system that is part of Westchester Center Health Network, also known as WMCHealth, was using a risk scoring algorithm in its electronic health record that was not very accurate.

THE PROBLEM

As a result, WMCHealth missed some high-risk patients and classified other patients as high-risk who were not. In addition, the automated daily report sent to case managers included only patients who had primary care doctors.

The case managers also wasted a lot of effort digging through charts to decide which patients to prioritize and which interventions to select. That reduced the amount of time they had to spend with patients.

PROPOSAL

The data management and analytics team at WMCHealth introduced business intelligence vendor Health Catalyst to the population health team at Bon Secours Charity Hospital.

“Health Catalyst is a vendor whose data operating system includes an open source machine learning package that the company has used to help other customers predict readmission risk with a high degree of accuracy,” said Deborah Viola, vice president of data management and analytics at WMCHealth.

Bon Secours wanted Health Catalyst to help the health system integrate the risk forecasts into its case managers’ workflow so they could easily use them in care management.

“We hoped that the case managers would like the application because it would increase their efficiency and allow them to devote more time to patients,” Viola said.

MARKETPLACE

There are many business intelligence vendors in the marketplace. They include Datawatch, Epic, Inovalon, Medhost, Periscope Data, Qlikview, SAS, Strata Decision Technology and Tableau.

MEETING THE CHALLENGE

To build the predictive model, Simer Sodhi and Lauren Torres of Bon Secours worked with Health Catalyst using the records of 54,000 patients who had been discharged from WMCHealth hospitals. This resulted in 24 risk factors that would be applied to the data in WMCHealth’s enterprise data warehouse.

“We wanted to validate the risk model’s predictions against particular patient cohorts and determine whether the algorithm would improve its accuracy as it learned from working with our data,” Viola explained.

“After we carefully validated the model’s predictions and we were sure that the risk predictions were more accurate than our EHR’s risk model, we added the risk scores to discharge lists on a new readmission risk platform that we integrated with our population health management registry software,” she added.

The advantage of this approach was that it leveraged a workflow the case managers were already familiar with, she added. They then paired the risk scores with EHR data and displayed the information on a dashboard that guided the case managers in identifying care opportunities and choosing interventions.

“The case managers used the discharge lists and risk scores to organize their work and prioritize the patients who needed to be engaged,” Viola said. “Having more accurate, more accessible data enabled them to follow up with at-risk patients faster. They also used this information in conversations with primary care doctors and specialists.”

As a result, they were able to obtain follow-up appointments faster – usually within seven days – and to connect patients with the services they needed to prevent emergency department visits and hospital readmissions.

RESULTS

The validation of the risk model showed a 17 percent increase in the number of discharges correctly classified as high- and low-risk. This was driven by an 8 percent increase in true positives (actual readmissions correctly classified as high-risk) and a 30 percent decrease in false positives (actual non-readmissions incorrectly classified as high-risk).

“The increase in accuracy was measured in relation to the LACE and EHR risk models,” said Simer Sodhi, director of data management and analytics at WMCHealth. “Seventy-two percent of the machine learning model’s predictions were accurate, compared to 62 percent of the LACE scores and 61 percent of the EHR scores.”

The difference was attributed to two factors: The machine learning model used many more variables than the others did, and it better reflected the characteristics of the health system’s population, she added.

Since the new discharge platform was introduced, it has cut the amount of time that the case managers need to prepare for their daily work from an average of 35 minutes to eight minutes.

“Incorporating the information on patient care gaps in the discharge report has also saved time,” Sodhi said. “Previously, our case managers had to spend two to five minutes per patient to identify these care opportunities. That added up to about 73 hours per year.”

Including work prioritization and care gap identification, the case managers are saving a total of 1,327 hours per year. As a result, the case managers have had more time to spend with patients and have been able to take on more patients who need their assistance.

ADVICE FOR OTHERS

Machine learning tools should be applied to the data on a health system’s patient population to increase the accuracy of readmission risk scores, Sodhi said.

“The risk scores should be part of discharge lists that are provided to case managers in their existing workflows,” she advised. “Supporting information from the EHR should be automatically paired with risk scores to make them more actionable.”

Finally, Sodhi said, case managers and frontline clinicians should have input into the development of the discharge platform so they can provide feedback on how well the platform meets their needs and fits with their workflows.

The post Machine learning system saves case managers 1,327 hours per year appeared first on Artificial Intelligence.