EVIDENCE BASEDINSIGHT
The Aspirational Role of Artificial Intelligence and Advanced Analytics
September 21, 2020 Ģý 6 min read
Ģý
Ģý is committed to exploring and implementing a fullrange of advanced analytic solutions that support clinical decisions. To ensurethat every new AI model will add real value to, the review processincludes a clinical pilot phase and an operational evaluation. This rigorousrequirement is a distinguishing characteristic of the Ģýapproach. As the company leverages its global expertise and vast data toaccelerate AI solutions for nephrology, itĢýs critical for researchers, data scientists,and practitioners to collaborate and prioritize tools that can be meaningfullyintegrated into a clinicianĢýs workflow.
Artificial intelligence (AI) is transforming healthcare throughadvancements in clinical decision support. AI is an umbrellaterm that brings together concepts from several fields such ascomputer science, statistics, algorithmics, machine learning (ML),information retrieval, and data science at large. Machine learningis growing rapidly due to advancements in computational powerand improvements in statistical techniques. Deep learning is asubfield of ML that mimics artificial neural networks to learn.These advancements ultimately enable AI to be leveraged toidentify hidden interactions and patterns within large, complex,multilevel datasets (Figure 1).
FIGURE 1|Relationship of artificial intelligence, machine learning, and deep learning
AIĢýspecifically machine learningĢýcan be defined as solvingproblems by building algorithms that are based on learning froma data set of examples utilizing stochastic probability-basedmodeling.1Implementation of AI in healthcare generally consistsof a series of stages as shown in Figure 2.
FIGURE 2|Application of AI in healthcare and kidney disease
In the problem definition stage, the clinical problem is definedand transformed into a relevant machine learning problem.For example, Ģý has a predictive modelto identify which patients are more likely to be admitted to thehospital in the next week. In the data preparation stage, largevolumes of retrospective data are collected, integrated, cleaned,and then engineered into variables (ĢýfeaturesĢý) for the model.In the hospital admission predictive model, clinical data (e.g.,hospitalization history, treatment vitals, laboratory measurements,nursesĢý notes) from over 100,000 patients is used to separate thosewho have been historically admitted or not admitted. In the model building stage, predictive models are trained and performance isanalyzed. Once the model performance is acceptable and solvesthe business problem, the model can then be deployed, used, andmonitored in production, the last stage. In the next section, we willreview some of the models that have been trained in healthcare.
In nephrology, while the application of AI is relatively limited,the potential is significant. A few nephrology-specific examplesinclude predicting, chronic kidney diseaseprogression, and morbid events in dialysis patients.2,3,4However,AI is ripe for nephrology due to the presence of large data sets andprevalence of disease. Clinical data of many patients treated withdialysis is stored in electronic medical records, making it ideallysuitable for AI applications. Clinical data for dialysis patients isunique (as compared to those in the hospital setting) because mostof the patients visit dialysis clinics three times a week for threeto four hours at a time. There are enormous amounts of clinicaldataĢýsuch as vitals, laboratory measurements, and comprehensiveassessmentsĢýcollected at each visit. Given this rich, diverse dataand access to large patient populations where these efforts can beproperly implemented, AI and other data-driven solutions can beutilized to improve patient care.
FRESENIUS MEDICAL CAREĢýS APPROACH
While AI has received a lot of attention and is a frequentlysearched topic online, it is one of several approaches to Ģýadvancedanalytics.Ģý5Typically, AI models are stochastic and require thecollection of lots of historic data to be able to predict the future.6But there are situations where historic data does not exist; if wedo not have data on how patients respond to certain medications,there would be no way to create AI models to describe patientresponse. In those cases, other approaches must be applied.
ĢýDeterministicĢý models can assist by building mathematicalmodels of physiological processes. Those types of modelingapproaches are part of a broader concept of advanced analytics.7A more holistic integrated approach to advanced analytics thatbrings together stochastic and deterministic modeling can beused with greater efficacy and a broader set of applications. Forexample, at Ģý, a team of mathematiciansutilized mathematical models along with stochastic elements toimprove the standard anemia algorithm for the companyĢýs dialysispatients. In some situations where comprehensive understandingof the physiology does not exist, AI is the best approach, suchas AI-based hospitalization predictive models used in FreseniusMedical Care Europe, Middle East, and Africa (EMEA), LatinAmerica (LA), and North America.
Ģý values incorporating AI into the fabricof the global organization. The first hospitalization predictivemodel created in 2010 by the Renal Research Institute was pilotedthrough a small Dialysis Hospitalization Reduction initiativewith a team of social workers.8Over time, with other efforts andmore internal expertise gained, analytics teams at FMCNA andĢý EMEA have been refining the process forapproaching advanced analytics.
In North America, AI activities are organized in a six-step qualityimprovement process: model creation, model prioritization,clinical pilot development, operational deployment, and outcomesanalysis (Figure 3). Model prioritization and a governing bodythat oversees the activities of the advanced analytics initiativeswere created in 2015. Ģý differentiatesitself by ensuring that the models go through clinical pilots andoperational deployment when necessary to ascertain their value.
FIGURE 3|FMCNAĢýs approach to advanced analytics
To date, the FMCNA steering committee has reviewed and ranked35 proposals. In December 2019, a global Ģýgroup of data scientists and mathematicians held a forum specificto Fresenius SE and Ģý. It brought togetherresources and expertise from various parts of the organization,allowing for more resource sharing, technical knowledge exchange,and collaboration to leverage global expertise.
Currently, nearly 40 advanced analytics algorithms have beencreated in FMCNA, and more than 70 efforts exist acrossFresenius SE companies. Successful implementation of AIsolutions requires diverse skill sets and the work of data scientists,statisticians, mathematicians, data engineers, and data analysts aswell as the full engagement of clinical leadership.
The Ģý experience with AI solutions suggeststhat the AI model development process may be analogous tothe drug development phases employed by pharmaceuticalcompanies. The initial aim is to understand whether the modelis predictive and whether the internal customers can use themin a pilot project. Latter stages focus on scaling to broader usewith interventions demonstrating whether the model makes adifference in patient outcomes (Figure 4).
FIGURE 4|FMCNAĢýs phased approach to AI model implementation
Some company AI solutions assist clinicians with understandingwhich in-center dialysis patients are at higher risk of fluid-relatedhospital admissions in the next week, while other solutions aimto identify patients at risk of developing foot ulcers. Other modelsseek to predict which in-center hemodialysis patients may begood candidates for. Effort is under way to bringtogether deidentified data from PD Cyclers, patient complaints,and dialysis electronic health records to understand whichperitoneal dialysis patients are at risk for leaving home modalities.
Some of these activities are deployed in small settings with onlya few users, while othersĢýsuch as the model that identifies besthome-therapy candidatesĢýare used across all FMCNA dialysisclinics. Implementation and operational workflow become animportant component of how models are utilized in clinical practice.
RISKS AND OPPORTUNITIES
AI and other advanced analytical activities have some inherentrisks.9,10,11One common concern is the bias that models canintroduce to patient selection. For example, speech recognitiontechniques for neurological diseases failed to account for patientlanguage differences and were improperly misdiagnosing patientswho spoke a different language.12Another concern is that patientsof certain ethnic or gender differences may be misclassifiedbecause they were not exposed to a certain treatment algorithm inthe past. Another common concern about ML algorithms is thatmany of them are Ģýblack boxĢý algorithms where the decision ofwhy patients are being classified as high risk may be unclear.13
Learning lessons from prior experiences should dictate how tobuild AI solutions in the future. It is important that the modelsare carefully created, piloted, and implemented with properdiscipline and clinician involvement, or else their many benefitswill be lost. Eighty percent of corporate executives report seeingpositive results from AI implementation.14Ultimately, with growingamounts of data available, AI solutions can help discern patterns nototherwise visible to clinicians. These solutions can often be quicklyand seamlessly brought into the cliniciansĢý workflow (Figure 5).
FIGURE 5|Risks and opportunities
Now is a unique time when vastly growing data, improvinghealthcare technology, and an abundance of scientific methods touse the data can be utilized to truly transform healthcare delivery.ĢýĢýs global outreach combined with dialysistherapy that yields vast amounts of data presents an exceptionalopportunity to utilize AI and other modeling efforts to improvepatient care. AI solutions made available to physicians and otherclinical support staff, as well as to patients, should be viewedas another clinical decision support tool to extend providersĢýinsights. AI cannot and should not replace providersĢý medicaldecisions, but instead assist them in providing optimal care. EveryAI effort should be carefully structured and monitored to assureethical and accountable implementation to provide the mostefficient and best patient care.
Meet The Experts
References
- Burkov A.The Hundred-Page Machine Learning Book.Self-published, 2019.
- Tomasev N, Glorot X, Rae JW, et al. A clinically applicable approach to continuous prediction of future acute kidney injury.Nature2019 Jul;572:116-19.
- Tangri N, Grams ME, Levey AS, et al. Multinational assessment of accuracy of equations for predicting risk of kidney failure: a meta-analysis.JAMA2016 Jan;315(2):164-74.
- Jiao Y, Chaudhuri S, Brzozowski J, et al. Predictive modeling is associated with reductions in hospital admissions in dialysis patients. Conference paper for National Kidney Foundation Spring Clinical Meeting.Am J Kidney Dis2015 Apr;65.
- Artificial intelligence in the news. Google. https://www.google.com/search?q=artificial+intelligence&source=lnms&tbm=nws&sa=X&ved=2ahUKEwjX9uqptO3mAhVFT98 KHdt7DiwQ_AUoAXoECBEQAw&biw=1536&bih=722. Accessed Jan 5, 2020.
- Stochastic modeling.Wikipedia. https://en.wikipedia.org/wiki/Stochastic. Accessed Jan 5, 2020.
- Deterministic versus stochastic models. From PowerPoint lecture on ĢýUnbounded population growth.Ģý https://www4.stat.ncsu.edu/~gross/BIO560%20webpage/slides/Jan102013.pdf. Accessed Jan 5, 2020.
- Jiao et al. Predictive modeling.
- Cheatham B, Javanmardian K, and Samandari H. Confronting the risks of artificial intelligence.McKinsey QuarterlyApr 2019. https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/confronting-the-risks-of-artificial-intelligence.
- Gruson, D. [The ethical risks associated with artificial intelligence must be identified and regulated.]Soins2019 Sep;64(838):48-50.
- Szabo, L. Artificial intelligence is rushing into patient careĢýand could raise risks.Scientific American, Dec 24, 2019. https://www.scientificamerican.com/article/artificial-intelligence-is-rushing-into-patient-care-and-could-raise-risks/.
- Gershgorn, D. If AI is going to be the worldĢýs doctor, it needs better textbooks.Quartz, Sep 6, 2018. https://qz.com/1367177/if-ai-is-going-to-be-the-worlds-doctorit-needs-better-textbooks/.
- London, AJ. Artificial intelligence and black-box medical decisions: accuracy versus explainability.The Hastings Center Report2019 Jan-Feb 49(1):15-21.
- Cheatham et al. Confronting the risks of artificial intelligence.