Impact of virtual triage and care referral on patient care seeking intent and clinical acuity alignment in an Australian health plan: A cross-sectional study

To evaluate if AI-based VTCR can improve care acuity alignment and patient experience while potentially reducing unwarranted clinical care services and avoidable care costs when implemented by a large Australian private health insurance company, Newcastle Industrial Benefits (NIB).

A cross-sectional study compared patient care intent before and following VTCR. NIB is a leading Australian private health insurance company insurance plans for Australian residents and international students and workers. An estimated 55% of Australian residents are covered by private health insurance plans, due to taxation benefits, reduced wait times for treatment, and access to ancillary healthcare services. As of 2024, NIB held an 9.8% share of the Australian private health insurance market.^[14]

The Infermedica VTCR engine conducts evidence-driven analyses spanning 800 illnesses, 1,500 symptoms, and 300 risk factors. VTCR can be accessed 24/7/365 in 24 different languages from any internet connected device. Using AI, machine learning, and natural language processing, the virtual triage engine assesses patient-users’ reported symptoms and medical history, and then identifies the most likely ailment and recommends the most suitable care setting based on clinical acuity or urgency. No predetermined protocols or decision trees are employed. With new information, the VTCR engine iteratively evaluates varied clinical hypotheses, just as a live clinician does. The data used in this analysis was extracted from VTCR encounters with patients using the Infermedica VTCR engine.

To ensure patient safety and minimize potential mistriage, AI-based VTCR engines require validation. The Infermedica VTCR engine will, by design, err on the side of overtriage to higher acuity care, rather than possibly mis-triage or under-triage a patient whose clinical presentation warrants higher acuity care. Across clinical specialties VTCR accuracy varies as a result of the differing depth and breadth of disease-specific data used to train the engine’s AI. Thus far, no instances of patient harm occurring as a result of virtual triage and care referral to an inappropriate level of care acuity have been documented. Clinical vignettes prepared by physicians in various clinical settings have evaluated the clinical validity and accuracy of VTCR.^[15–17] Prior research demonstrated that the Infermedica VTCR engine provides safe recommendations in 97.8% of cases and is equal to or superior in accuracy to rules-based triage protocols.^{[15, 17]}

Analyses focused on international plan members (students and workers) resident in Australia, with data collected from March 1st to October 31st, 2024. During this period, 8,195 encounters were completed, of which care seeking intent was determined for 4,471 encounters. Post-VTCR intent was inferred from a combination of intent survey patient response and patients engaging the call-to-action to book an appointment within the application. Patients also reported on the quality of their VTCR user experience. Patients provided explicit consent prior to the virtual triage encounter for their data to be analyzed in a fully de-identified manner and presented in the aggregate.

Patient-user data was examined to determine the level of clinical care acuity alignment or divergence between patient post-triage care intent and that recommended by VTCR. The care intent survey provided the following care options: (1) self-care for symptoms not requiring professional medical care, managed by patients at home; (2) outpatient medical consultation, where symptoms warranted medical evaluation by scheduling a routine in-person visit with a healthcare provider; (3) outpatient consultation within 24 hours, where symptoms indicated a more urgent need for consultation with a healthcare professional; (4) emergency department (ED) care, where symptoms were sufficiently serious that patients were recommended to proceed immediately to the nearest ED or to call an ambulance; and (5) patient did not know or were uncertain about what level of care to seek. Patient care intentions before and after virtual triage were analyzed and differences were tested for statistical significance.

The findings are presented by identifying the volume of patients per 1,000 completed encounters at each of the five levels of clinical acuity. Potential financial return on investment (ROI) was estimated per completed interview, based on the following mean claims costs per visit: ED visit cost of $394; outpatient primary care general practitioner visit cost of $52; and $49 for a telemedicine consultation. Estimated potential savings are presented using three measures: (1) potential net cost savings per VTCR encounter across all levels of care acuity post-triage; (2) potential net cost savings per encounter involving a de-escalation of care acuity from pre-triage emergency care intent to a lower acuity level; and (3) potential net savings per interview based on de-escalation from pre-triage outpatient care intent to self-care.

Patient self-reported disease risk factors, relevant for engagement of health promotion and risk reduction programs offered by the health insurer, were collected during VTCR encounters and reported.

As all information collected for these analyses was deidentified and anonymized, and all patient information is reported only in the aggregate, there is no requirement in Australia for consent as there is no personal information involved or risk of patient informational harm. Based on this, NIB was comfortable with the use of the de-identified data for research purposes as low risk and ethics review board approval was not sought.

There were 8,195 completed VTCR encounters between March 1 and December 31, 2024, and in 4,471 of encounters the pre- and post-triage intent surveys or call-to-action were completed by patients. Regarding cohort demographics, 64.5% of patients were female and 97.2% were aged 18-44; 65.8% were 18-29 years old and 31.4% were 30-44 years old. Almost all patients chose to complete the VTCR encounter in English.

There were 353 actionable risk factors reported per 1,000 encounters, including individuals with obesity/overweight status, diagnosed hypertension and hypercholesterolemia, and/or who used tobacco products, highlighting the opportunity to recruit patients for health promotion and disease risk reduction programs through VTCR.

Patients reported high satisfaction with the VTCR experience, conveying a mean rating score of 4.2 on a 5-point scale, with 39.8% evidencing uncertainty in healthcare seeking which decreased following use of VTCR.

As seen in Table 1, the overall level of compliance or alignment with triage recommendations, as indicated by reported care intent, was high (74.0%), demonstrating that VTCR was effective in educating patients about the most clinically appropriate care to meet their needs. One-half of members (50.5%) changed their reported care seeking intent as a result of the triage recommendation. Among those who changed care seeking intent, 7.3% (or 3.7% of all members) de-escalated their intent to lower acuity care, compared with 11.6% (or 5.8% of all members) who escalated their intent to higher acuity care.

As can be seen in Table 2, following the VTCR encounter, there was a 91.3% decrease in patients with uncertain care seeking intent, a reduction of 39.8 percentage points (PP) from an absolute value of 43.6% of patients pre-triage to 3.8% of patients post-triage. There was a 56.5% (6.2 PP) increase intent to engage self-care following VTCR, benefiting the patient, health system and health insurer alike by reducing unnecessary and avoidable care visits. A 35.7% (0.5 PP) decrease in emergency care occurred following virtual triage, from 1.3% to 0.8% of patients. All of these post-VTCR changes in patient care intent were statistically significant (p < .05), and can achieve potential reductions in avoidable care and associated costs across the patient population.

As shown in Table 3, the volume of patients selecting routine outpatient care increased substantially following virtual triage and automated care referral. Based on the change in patient pre- and post-virtual triage care intent, there was a potential $4.27 (8.6%) total net saving per completed VTCR encounter.

Table 4 shows that nine unnecessary ED visits were avoided per 1,000 VTCR encounters, yielding estimated potential net savings of $284.55 (72.2%) per completed encounter among patients with an initial intent to seek emergency care.

Table 5 shows that 35 unnecessary outpatient visits were avoided per 1,000 encounters, yielding a potential estimated savings of $3.39 (6.5%) per completed encounter among patients who had an initial intent to seek outpatient care.

Virtual triage and care referral improved acuity alignment between patient pre-triage care intent and actual clinical need, with a 91.3% reduction of patients with indecision or uncertainty about what care to seek. This may yield a net reduction in avoidable medical costs across a health plan member population by engaging patient care earlier in the course of disease, when therapeutic interventions are often less complex and costly. Improved care acuity alignment can reduce care delays, which often result in greater hospital length of stay and potentially avoidable ICU admissions.^{[8, 19–22]} De-escalation of care acuity was particularly notable in a large increase in intent to engage self-care following VTCR (56.5%, 6.2 PP), which can potentially reduce avoidable or unnecessary care service delivery.

The greatest post-VTCR escalation in care acuity occurred in routine outpatient care, which almost doubled in absolute magnitude (and 37.2 PP). This indicates an ability of VTCR to expedite needed care while patient acuity is still low, when care effectiveness is high and costs and patient dislocation are low relative to urgent outpatient (within 24 h) or ED care. While most patients with symptoms acute enough to warrant consideration of an ED visit will not consult with VTCR, there was nonetheless a concomitant decrease in ED care intent by 35.7% (0.5 PP). These changes in care recommendation were regarded by patients positively, as indicated by an 84% VTCR satisfaction rate. These findings are similar to those reported in other analyses.^{[10, 12]} Also notable in the overall use pattern of VTCR was the technology’s ability to encourage candid patient reporting of disease risk factors, with 35.3% of encounters where such data was reported.

These findings parallel other studies which evidence that VTCR, through its ability to appropriately realign patient care intent with the level of care acuity actually needed, can deliver meaningful clinical and operational value by reducing avoidable, clinically unwarranted higher acuity ED and urgent outpatient care utilization.^{[10, 12]} We estimated the potential net cost savings per encounter in this cohort of patients was $4.27 or 8.6% across all levels of care acuity, and $288.55 potential net savings (72.2%) for de-escalation of care from unneeded ED visits to routine, non-urgent outpatient care. Further, 35 unnecessary outpatient visits were avoided per 1,000 encounters, conveying estimated potential plan savings of $3.39 or 6.5% per completed encounter among patients who intended to seek outpatient care prior to VTCR. In addition, almost 40% or 399 patients per 1,000 encounters intended to book a telemedicine visit as recommended after virtual triage. These estimates do not consider the patient and clinician satisfaction resulting from expedited respective care flows and lower wait times for patients to receive care, qualitative outcomes that should also be the focus of further research. We are also unable from the data to estimate the value of clinical care capacity created within care delivery workflows by relieving clinicians from delivery of care to patients whose clinical acuity does not warrant an in-person consultation.

In interpreting these findings, it is important to bear in mind that the final care acuity decision of patients may often be influenced or limited by their current health plan coverage, access to resources for self-care, and the availability of and ability to access non-urgent outpatient care services. Patient sampling may also introduce potential selection bias, as patients engaging with digital health tools may be systematically different in key characteristics from the general population in terms of internet access and literacy, health literacy, or care engagement and seeking behavior. VTCR users in this cohort of patients were disproportionately younger and female, which limits the generalizability of these findings. For example, elderly patients may experience anxiety or have fears about using technology, and may be less candid in disclosing key health or symptom information during the VTCR encounter.

This study design did not allow for the collection of data regarding possible health disparities based on patient income, race, or those with literacy challenges, or who may have marginal access to technology. Future research should endeavor to capture pertinent data on these important variables that impact healthcare intent, care seeking behavior, ability to secure care, and patient outcomes.

The current study design also relied on self-reported healthcare intent, which may not fully translate into actual care seeking behavior or care services delivered. Reliance on patient self-reporting of care intent prior to and following VTCR as a means to assess the impact of the technology on patient healthcare choices and care seeking behavior is a limitation of the present study. Patients may not always be truthful in reporting their care intentions during VTCR encounters, and actual patient care seeking behavior following VTCR might not align with the intent reported during the triage encounter. Post-VTCR validation of actual care sought in future research can evaluate whether or not reported patient intent translated into real world care behavior, reducing the potential issues around the internal validity and generalizability of the findings reported in the present study.

It is of course impossible to precisely quantify the financial value of specific changes in care acuity realignment without data conveying actual patient care flow and utilization, site of intake, services rendered and clinical outcomes - all beyond the scope of this study. Future research should include clinical validation of patient diagnosis and care services delivered as a primary outcome in order to confirm the potential savings identified in this study. Nonetheless, the combined impact of all changes observed in care seeking intent was a substantial overall potential reduction of inappropriate care utilization and associated care and administrative costs, as well as potential recovery of clinical care capacity to accelerate appropriate treatment of higher acuity patients with more urgent care needs. This likely improves clinical outcomes and financial performance, although further research is needed to validate the full impact of improving care acuity alignment through VTCR.

Systematically and prospectively evaluating how changes in care intent impact care utilization and health outcomes should constitute the next phase of research on the integration of VT into existing triage and care referral workflows. Given advances in generative and conversational AI, most recently the advent of large language models, VTCR may evolve to offer a voice automated capability. Voice automation in VTCR may convey further reductions in the costs of live and automated call centers, and can divert routine cases from busy clinicians so they have more resources to focus on higher acuity, more urgent patient presentations. VTCR impact on the clinical and financial performance of health plans and healthcare delivery organizations should also be an ongoing and expanded focus of future research.