Machine Learning as an Enabler for Augmented Intelligence

By Kapila Monga, MBA

There is a proverb that says, “When the winds of change blow, some people build walls, and others build windmills.” In truth, the ubiquitous emphasis on adoption of artificial intelligence (AI) and machine learning (ML) today is a bit more than just a wind of change. And, for the healthcare industry in particular—because of ongoing regulatory changes, privacy considerations, and biases and inequities in the data—widespread and accelerated adoption of AI/ML has been somewhat challenging. This does not undermine the successful use of AI in applications like reducing wait times in hospitals, patient triaging, and shortening intensive care unit stays, but the industry at large has yet to see significant benefits from this new technology. Thus, it is worthwhile to take a step back and identify an acceptable way for the healthcare industry to be able to use AI and ML at scale.

Augmented intelligence is a subsection of AI/ML developed to enhance human intelligence rather than operate independently of it or replace it. It is designed to improve human decision-making and, by extension, actions taken in response to those improved decisions.

Augmented intelligence might prove to be the key that the healthcare industry needs to harness the power of AI.

  • With augmented intelligence, humans retain the final decision-making authority in tasks of their choosing.
  • The algorithm can be designed to operate under selective conditions, helping to avoid some of the data biases and health inequities related challenges.
  • Shutting off/temporarily pausing the algorithm at any point is easier than a typical general AI solution.
  • Explainability and transparency can be made to be mandatory tenets within the design of the augmented intelligence solution.

While a lot has been written about potential use cases of augmented intelligence in healthcare industry overall, a framework on how to decide if a use case from that list is a good candidate for augmented intelligence (within the context of an organization) is warranted.

Identification of the Right Process to be Augmented by Augmented Intelligence

Identifying the right process to be augmented by augmented intelligence is a critical activity. To determine whether a process is a good candidate, assess the extent to which the following questions can be answered succinctly by the process experts:

  1. Quantification of baseline metrics: Are the current efficiency and accuracy levels of the process (without augmented intelligence in place) well understood (i.e., quantified)? The accuracy metric will be dependent on the process in question, but there should be a business consensus on the definition of the accuracy metric, and a number associated with the metric should be available and documented. Populating the table below should be a requisite:
Process nameCurrent turnaround time / efficiencyAccuracyData conditions / boundary conditions
<Process Name>

e.g.: Documenting missing diagnosis codes based on chart review

<Current speed at which process is happening>

e.g.: 16 requests per day per analyst

<Accuracy of current process>

e.g.: Accuracy metric: # of codes accepted by physicians on review

Metric value: 75 %

<On what data universe the time and accuracy metrics have been calculated>

e.g.: For data from the state of New Jersey; for new patients with age >45 and <=65, etc.

  1. In-principle alignment on scope of improvement in the process: Based on the current process understanding, is there a scope of improvement either in terms of speed or accuracy for the current process? In other words, if process experts are interviewed, will they be able to respond affirmatively on the existence of the scope of improvement, and will they be able to point to specific factors/drivers that make them conclude the same? Being able to point to specific instances/examples or test cases that demonstrate the existence of the scope of improvement is also sufficient. However, just believing that there is a scope without having anything to substantiate it, is not.
Questions for Process Subject Matter Experts
Is there room for improvement in accuracy and/or speed of the current process?

Illustrative Answer: Yes, there is scope of improvement in both accuracy and speed.

What are the drivers of the anticipated improvement in either the accuracy or the speed?

Illustrative Answer: Typically, when we are reviewing electronic health records for patients having COPD and diabetes, there are few diagnosis codes that are almost always missing. We end up adding them manually for almost everyone with these medical conditions.

Are there any instances you can share where you feel the process should have done better, and why do you think that is the case?

Illustrative Answer: Yes, we can pull out a report which will have chart IDs, and also have information on which Diagnosis codes were missing / were  added by us during review and why we added the same

  1. Benefit of an increase in a percentage point of accuracy or speed in the process in question: Quantification of the benefit in the right (read “dollar”) terms like cost savings or revenue increase will enable a cost-benefit analysis and ensure organization wide focus and concerted efforts for augmented intelligence.
  2. Alignment on desired level of transparency and explainability in the results: Will the outputs of the augmented intelligence process give you enough ammunition for validating the results? One key reason augmented intelligence can’t be adopted in all fields of healthcare is lack of desired level of explainability in results/transparency on how the results are calculated. At this point, it is worthwhile to get aligned on the kind of outputs expected from augmented intelligence processes so that both developers and users are convinced of the usability. In absence of this conviction, the process shouldn’t be considered as a candidate for augmented intelligence.
Augmented Intelligence in Action

As an illustration, let’s explore how augmented intelligence would work for the use case of documenting missing diagnosis codes based on medical chart review.

  1. Training an ML algorithm to identify missing codes: This is the first step. The ML algorithm in this case will identify the missing codes after doing the chart reviews using natural language processing (NLP)-based methods. The main data source for this step would be data from EHR systems and rules/business knowledge that teams have been using so far for identifying diagnosis codes from chart reviews/chases. At the end of this step, teams will get an NLP algorithm that will generate the following type of outputs for each EHR record it processes:
Patient IDChart IDMissing CodeEvidence from Chart Data
PATID_123CHT_123C34.30 – Malignant neoplasm of lower lobe, unspecified bronchus or lung–       Family history

–       Problem list has codes for neoplasms and lung infection

–       Genetic test for bio-marker xyz was positive

  1. Compare the algorithm accuracy against the current process: In this case, both the algorithm and the current processes should be made to process the same set of records in parallel, and statistics on time taken should be noted. It is worth noting that since we are talking about augmented intelligence, validating the output and adding it to the chart will still continue to happen the way it was before augmented intelligence. So, in essence, this algorithm is only augmenting the work of HI professionals, and not replacing it. Final decision on whether to consider or act on the output of the machine learning algorithm still rests with the HI professional. Results post-manual validation by the HI professional should also be noted. A decision then needs to be made on whether the algorithm needs further refinement, comparing its speed and accuracy against that of current way of doing things. It is recommended to run the machine process and human process in parallel, and refine the machine process until it is as at least as accurate as the human process and the speed is much faster than the current human process.
  2. Augmented intelligence in live action: Once all stakeholders are satisfied with the current level of accuracy, the machine learning process around missing code identification should be set as Step 1 of the overall HI professional process flow. In Step 1, HI professional gets the results from the algorithm (i.e., missing diagnosis codes along with evidence). In Step 2, the HI professional processes the result by either adding the missing code to the chart or decides that the evidence isn’t sufficient enough to add the code and passes this feedback back to the ML model. This feedback is then used to refine the ML model so that iteratively both accuracy and speed of the ML model become better . In essence, the ML algorithm in this case is augmenting the intelligence already possessed by the HI professional, and the HI professional still remains the driver of the overall decision-making. It is worth mentioning here, that there could be processes where entire decision making can be automated as well. However, even in those cases augmented intelligence should be the first step. After having deployed and used augmented intelligence in the process for at least a couple of end-to-end process cycles, a data-driven determination should be made of whether complete automation is a wise decision or not.

Not all processes in the healthcare industry are ready for AI, but most can benefit from augmented intelligence. However, businesses must still perform a cost-benefit analysis for using augmented intelligence. In the event that a process is suited neither for AI nor for augmented intelligence, it may be wiser to deliberate on other strategies to improve the process efficiency and effectiveness. A few examples of such strategies include merging the process with an upstream/downstream process, implementing behavioral changes, and offshoring or outsourcing. The underlying objective of all technological innovations is to help businesses operate faster and better; when this objective can be achieved, means should be of secondary importance.

Kapila Monga ([email protected]) is a director in Cognizant’s Digital Business AI and Analytics Practice.

Leave a commentSyndicated from

Translate »
%d bloggers like this: