>> Chronic hepatitis C virus (HCV) infections are the most common cause of liver cirrhosis, liver cancer, liver transplants and death from liver disease. The main risk factors for becoming infected with HCV are intravenous drugs, blood transfusions (before 1992), unprotected sexual intercourse with changing sexual partners, hemodialysis or tattoos if studios ignore basic hygiene rules.
The Robert Koch Institute, Berlin, reports a significant increase in reported new infections between 2000 (5,091 patients) and 2019 (6,633 patients), with an unknown number of unreported cases. As a result, the World Health Organization's (WHO) declared goal of eliminating viral hepatitis in all countries by 2030 is becoming a distant prospect. New strategies are required, such as the use of algorithms from the field of artificial intelligence (AI).
All HCV patients should be protected from
benefit from innovative therapies
Background: From a medical perspective, the high number of unreported cases of HCV infections is particularly annoying because there are now innovative therapies. That was not always so. For a long time, hepatologists only had the option of treating HCV with pegylated interferon α plus ribavirin. This regimen was not always successful, but was associated with a number of side effects.
Direct-acting antiviral agents, i.e. directly effective antiviral drugs, are now available. They inhibit viral proteases or polymerases. Depending on the HCV genotype and drug, doctors can cure 90% or more of all patients - provided they recognize the disease in time.
“Reaping the potential benefits of novel treatments requires reducing the size of the undiagnosed population and establishing early diagnosis so that patients can be treated before long-term sequelae of HCV infection occur,” write Drs. Orla M. Doyle and colleagues in Scientific Reports. You work at IQVIA in the area of Predictive Analytics, Real World Solutions.
Innovative approaches to
Identify patients at risk
The problem of high rates of patients with undetected HCV is not new. The Centers for Disease Control and Prevention (CDC) in Atlanta introduced a risk-based screening program years ago. Their target group was patients who used intravenous medications or had certain diagnoses such as HIV infection. There were also patients who had received blood transfusions before 1992. In 2012, the CDC expanded its recommendation to include people born between 1945 and 1965. 75 percent of all diagnosed HCV infections are found in this population. Nevertheless, the number of detected HCV infections has only increased marginally.
What to do? Doyle and colleagues see great opportunity in screening programs using artificial intelligence (AI) and machine learning tools. The basis is data from electronic patient records. No direct contact with patients is required initially.
It works like this: In the first step, you train an algorithm with patient data including information about whether HCV infections are present or not. In the second step, the performance of the program will be tested on an independent set of patient data for which information about HCV status was known but has been removed by the researchers.
If such algorithms are later used in practice, it will be possible to find patients with a high risk of an undetected HCV infection in a short time. Doctors only use them to carry out laboratory tests. However, you do not need to screen large numbers of patients with virological tests. In addition, AI can potentially reduce the number of false positive tests, i.e. people who are examined in the laboratory due to perceived risks but are HCV negative. This saves costs in the healthcare system.
Data sources for analysis
Doyle's team worked with existing longitudinal databases from IQVIA. These included the LRx database with medical prescriptions (Rx preparations) and the Dx database with medical diagnoses that were not verified by a third party. Entries came from the period between January 2001 and September 2016.
The LRx database includes two billion prescriptions per year, with coverage of up to 92 percent for community pharmacies, 70 percent for mail order and 70 percent for long-term care. The basis is electronic data that is anyway generated or processed by pharmacies, insurance companies, software providers and transaction clearinghouses. This includes information about the product, the provider, payment – private or as insurance reimbursement – and the patient's place of residence.
Dx data, in turn, contains fields on the patient's demographics, i.e. their age, gender, and also their place of residence. There are also details about payment of benefits via insurance or payments out of your own pocket. It is also noted whether patients were at the doctor's office or in the hospital.
All information remained anonymized throughout the data collection, linkage and analysis process. Personal health information was removed or encrypted before being collected by IQVIA. Both the LRx and Dx records are linked back to an anonymous patient token.
Selection of patients for the study
In the further course of the analysis, HCV patients were defined by a diagnostic code or by a specific treatment. The criterion for including additional therapies was important because a disease was not always mentioned in the data. In addition, there was the earliest time of a diagnosis or prescription. Patients without HCV had therefore received neither HCV diagnoses nor HCV medication.
Doyle's team worked with a case-control design. For each HCV patient, a patient without HCV but with comparable characteristics was selected. Through this matching, the scientists tried to ensure that differences in the distribution of predictors between HCV and non-HCV patients reflect medical phenomena and not just arise from different observation periods.
At the same time, the researchers worked on stratifying their data. They wanted to prevent their algorithm from only using general, well-known risk factors such as hepatologist care, intravenous drug use, alcohol abuse, other chronic viral infections such as HIV, but also kidney and respiratory diseases and other internal diseases as criteria to identify patients at risk. The CDC had previously failed with exactly this approach.
In addition to this stratification, all patients were required to meet the following eligibility criteria within the linked LRx and Dx data set: They should have an index date between October 2015 and September 2016. This refers to the first contact with the doctor for unclear symptoms. In addition, there was at least 24 months of activity, i.e. several doctor contacts during the period. Individuals were between 20 and 79 years of age on the index date and had medical visits or prescriptions registered in both databases.
At the same time, experts developed 98 clinical concepts to map associations between HCV infection and possible comorbidities, such as risk factors (intravenous drug use), symptoms (jaundice), treatments (for opioid use) and comorbidities (depression). To exploit the longitudinal nature of the data, features were represented using two metrics: frequency of occurrence and time of first occurrence.
Statistical methods
To ensure that model estimation, selection and evaluation were performed independently, the data were randomly divided into three sets: a training set (80 percent of HCV patients) for model development, a validation set (10% of HCV patients ) for model evaluation and selection, and a test set (10 percent of HCV patients) to assess the performance of the final model.
The researchers applied conventional parametric methods such as logistic regression (LR) to their data. There were also non-parametric machine learning methods such as Random Forest (RF), Gradient Boosting Tree (GBT) or Ensemble Learning based on GBT. These algorithms were chosen to integrate a classic approach from statistical modeling as well as two of the most popular tree-based machine learning approaches. Results from LR, RF and GBT were compared with other statistical methods.
Data analysis results
For the HCV cohort, 120,023 patients met the stratification and selection criteria. In the subgroup without HCV there were around 60 million, of which 9,601,900 were randomly selected.
Age and gender were in all
Cohorts similar. In the HCV cohort, 11% were treated with HCV drugs. A significantly higher proportion of HCV patients had a diagnosis of HIV or AIDS (5% of HCV patients versus 1% of non-HCV patients) and/or a history of intravenous drug use (26% of HCV patients versus 3% of non-HCV patients -patients).
The analysis of the data showed that, on average, patients already had symptoms two to three years before their actual HCV diagnosis.
symptoms such as joint pain, abdominal pain, malaise, fatigue or fibromyalgia. They received nonsteroidal anti-inflammatory drugs (NSAIDs), systemic steroids, or opioids three to four years before their HCV diagnosis. Patients contact the doctor about their complaints at an early stage. Several examinations follow, but a long time passes before the HCV diagnosis is made (Fig. 1).
Results of the predictive
Modeling
AI algorithms could shorten this period of time. But what do you want to achieve later in practice? If the goal is to find 10% or fewer of the HCV patients without a diagnosis, which the authors refer to as recall, all three models examined perform similarly. At recall levels above 30%, GBT produced the best results. At a value of 50%, for every 100 patients identified by the model, 74 were HCV positive. However, at a recall level of more than 50 percent, the precision decreased significantly.
The GBT model was then retrained using temporal features. Specifically, the precision increased from 74 to 87% at 50 percent recall.
Ensemble learning methods were applied to all models to train a stacked ensemble as a combination. It was then applied to the test data set. At higher recalls, the stacked ensemble outperformed all models. It achieved a specificity of 99.9% and a precision of 97% with a recall level of 50%.
That means: About 97 out of 100
Patients will be correctly identified as HCV if the goal is to find at least 50% of HCV patients.
The stacked ensemble works with 284 variables. In addition to age and the use of intravenous medications, painkillers, namely NSAIDs and opioids, proved to be particularly relevant for distinguishing between HCV and non-HCV patients. Treatments related to osteoarthritis and rheumatoid arthritis were also considered relevant (Fig. 2).
Limitations of the study
“The evidence in this study points to the significant potential benefit of AI in addressing problems of underdiagnosis associated with complex disease features,” the authors write. They point out four limitations:
• Only the history of HCV patients up to the point immediately before diagnosis or treatment was recorded. Patients could possibly benefit even more from an earlier start.
• In the first step, experts assessed and aggregated clinical characteristics. This hypothesis-driven approach can be flawed.
• An external prospective validation of the AI tools is still pending.
• Patients with limited access to the healthcare system were underrepresented in the sample: a problem that is certainly more relevant in the USA than in Germany.
• Some patients with HCV received neither diagnosis nor treatment during the period, which may have biased the results.
AI: New approaches for
Research and practice
“Artificial intelligence solutions have the potential to transform how HCV patients are identified and achieve the WHO goal of eliminating viral hepatitis by 2030,” wrote Doyle and her co-authors. You will see several options for deploying systems after evaluating them.
AI tools could help doctors identify people at high risk of HCV without a diagnosis in digital patient records, without even calling them into the office. Blood tests are only carried out if there is evidence of an increased risk. Such applications could run permanently in the background to constantly screen patient data - both in the hospital and in the doctor's office. Last but not least, the authors see possible uses in recruiting participants for clinical studies - be it to find or exclude patients at high risk without the need for laboratory tests in the first step.
“The prospect is undoubtedly exciting,” the article concludes. But there are many hurdles that must be overcome, from infrastructure to interoperability to information management and clinical implementation. At least the first step has been taken: AI algorithms are suitable for detecting patients with an increased risk of HCV in routinely collected health data. The researchers have shown this. <<

from: Dr. Gisela Maag, IQVIA Commercial GmbH & Co. OHG, Frankfurt