The Last Mile: Why Patient Adherence Determines Whether Early Detection Saves Lives

When a blood test signals danger and a patient does nothing, early detection fails. This paper examines the scale, causes, and solutions to the adherence crisis at every stage of the care journey — and proposes a dedicated REDI Last Mile Division to solve it.

Introduction: Detection Without Action Is Not Early Detection

The Ratner Early Detection Initiative (REDI) is building a suite of machine learning algorithms that read pre-diagnostic signals for 24 diseases — 13 cancers and 11 non-cancer conditions — from routine blood panels already drawn at annual physicals. The program can flag a patient as high-risk for lung cancer, colorectal cancer, ovarian cancer, or type 2 diabetes months before clinical diagnosis through any other means. A risk flag, however, is not treatment. A high-risk notification has value only if the patient who receives it acts on it. The space between notification and action is what this paper calls the last mile.

The last mile encompasses every step a patient must take after receiving a high-risk signal: scheduling and attending a specialist visit, completing confirmatory screening, accepting a diagnosis, initiating treatment, and returning for surveillance. It also encompasses a wider and older problem in medicine: the persistent failure of patients to take prescribed medications, keep follow-up appointments, or sustain behavioral changes that protect their health. These two challenges — the cancer-screening last mile and the general medication-adherence crisis — share the same root causes and respond to many of the same interventions. Understanding both is essential to designing an early detection program that actually saves lives.

This paper reviews the scope of the adherence problem, its documented causes, the state of the evidence on what works, the rapidly evolving role of AI-powered communication tools, and a concrete plan for a new division of REDI dedicated entirely to solving the last-mile problem for every patient flagged by REDI’s algorithms.

The Scale of the Adherence Problem in Chronic Disease Management

Medical non-adherence is one of the largest preventable causes of illness and death in the United States. Roughly 50 percent of Americans on long-term therapy for chronic disease do not take their medications as prescribed.¹ The consequences are staggering: approximately 125,000 deaths per year are attributable to medication non-adherence, along with up to 25 percent of all hospital admissions and an estimated $289 billion to $500 billion in avoidable healthcare costs annually.²

The burden is not evenly distributed. Patients managing cardiovascular disease who do not adhere to their medications experience dramatically worse outcomes. A 2024 global analysis found that a 20 percent improvement in cardiovascular medication adherence correlates with an 8 percent reduction in cardiovascular events and a 12 percent decrease in mortality.³ Medicare patients who missed adherence benchmarks for diabetes, hypertension, and hyperlipidemia faced 21 to 50 percent higher rates of inpatient stays and emergency department visits, along with 11 to 13 percent higher total healthcare costs, compared to adherent patients.⁴

Non-adherence takes several forms. Patients may never fill an initial prescription — so-called primary non-adherence. They may fill a prescription but take the medication inconsistently or stop before the course is complete. They may intentionally modify doses in response to side effects or cost concerns. Or they may simply forget. Each form is driven by a distinct set of barriers and responds differently to intervention.

Why Patients Do Not Adhere: A Taxonomy of Barriers

Research consistently organizes adherence barriers into four interacting domains: patient-related factors, therapy-related factors, healthcare system factors, and socioeconomic factors. In practice, barriers from multiple domains combine and reinforce each other.

Patient-Related Factors

The largest category of non-adherence predictors is psychological. Six of the eight most common predictors identified in a 2023 systematic review were rooted in patient psychology and health literacy.⁵ Patients who do not understand why a medication is necessary, who believe they feel fine without it, or who fear side effects they have read about online are statistically more likely to stop taking it. Nearly 40 percent of patients in one study reported not fully understanding their prescriptions when leaving their physician’s office.⁶

Mental health adds another layer of complexity. Depression and anxiety are associated with significantly lower adherence across virtually every chronic condition studied, creating a cycle: the disease drives psychological distress, which undermines adherence, which worsens the disease. Cognitive decline, particularly among older adults, poses its own challenge. Approximately 40 percent of older adults with mild cognitive impairment struggle with consistent medication management. Polypharmacy — the use of five or more medications simultaneously — doubles the likelihood of adherence difficulties.

Therapy-Related Factors

The design of the treatment regimen itself matters. Complex schedules, frequent dosing, and medications with prominent side effects are all independently associated with lower adherence. A 2022 systematic review found that patients who had previously experienced an adverse drug reaction were 50 percent more likely to be non-adherent to subsequent prescriptions.⁷

Healthcare System Factors

Inadequate patient-provider communication is a systemic adherence barrier. When physicians do not explain the rationale for a treatment, patients are less likely to follow through. Structural barriers — long waits for appointments, poor care coordination across specialists, absence of systematic follow-up — create friction at every step of the care journey.⁶ Most health systems still lack mechanisms to close the loop after an abnormal screening result or a new prescription is issued.

Socioeconomic Factors

Cost is among the most powerful adherence barriers. Prescription abandonment rates are below 5 percent when the drug carries no out-of-pocket cost; they rise to 45 percent when the copay exceeds $125 and to 60 percent when it exceeds $500.⁸ Among low-income patients, some face a literal trade-off between medication and food. Patients who sacrifice basic needs to pay for medication represent one of the highest-risk groups for non-adherence.⁹

The Last Mile in Cancer: From Risk Flag to Resolved Care

In the context of REDI’s early detection algorithms, the last mile is not a single step. It is a sequential journey that a high-risk patient must complete successfully at every stage. A failure at any point forfeits the survival benefit that early detection would otherwise confer.

The Journey Defined

When REDI’s algorithms generate a high-risk signal — for example, a CBC pattern consistent with pre-diagnostic lung cancer — the patient must take a specific series of steps to convert that signal into a survival benefit:

Receive and understand the risk notification from their primary care physician

Schedule and attend a specialist referral or confirmatory imaging appointment

Complete confirmatory diagnostic testing (e.g., low-dose CT scan for lung, colonoscopy for colorectal)

Accept a diagnosis if one is confirmed

Initiate definitive treatment without delay (surgery, chemotherapy, radiation, or combination)

Return for surveillance at prescribed intervals to detect recurrence

Each of these steps is an opportunity for dropout. The clinical literature documents high failure rates at multiple points along this chain, even in the existing standard-of-care screening programs.

The Follow-Up Failure After Abnormal Screening Results

The failure to follow up after an abnormal screening result is one of the most thoroughly documented problems in cancer medicine. A 2023 NCI-funded clinical trial of nearly 12,000 patients found that fewer than one quarter of patients overdue for follow-up after an abnormal result on a screening test for breast, lung, cervical, or colorectal cancer completed that follow-up testing within four months under usual care.¹⁰ Even the most intensive intervention tested — an EHR reminder combined with a letter and a phone call — raised the completion rate to only about one-third.

For colorectal cancer specifically, the gap between stool-based screening and colonoscopy follow-up is especially concerning. While landmark clinical trials achieved follow-up colonoscopy rates of 80 to 90 percent among stool-test-positive patients, real-world safety-net data tell a different story: only 18 to 56 percent of patients with a positive stool-based test completed follow-up colonoscopy within one year.¹¹ Patients who never complete follow-up colonoscopy after a positive fecal immunochemical test face a two-fold increased risk of dying from colorectal cancer compared to those who do complete the test.¹²

In many health systems, no mechanism exists to actively track patients who fall out of the follow-up pipeline. Accredited facilities billing Medicare for lung cancer screening must submit data to a registry, but — as a 2021 NEJM commentary noted — there is no requirement that the registry be used to close the loop and ensure follow-up actually occurs.¹³

Stage Matters: The Survival Stakes of the Last Mile

For the cancers in the REDI portfolio, the survival difference between Stage I and Stage IV is measured in decades, not percentages. Colorectal cancer diagnosed at the localized stage carries a five-year relative survival rate of nearly 90 percent. Diagnosed at the metastatic stage, the five-year survival rate falls to approximately 13 percent.¹⁴ A patient whose CBC pattern is flagged as high-risk, who receives a confirmatory colonoscopy, and who has an early-stage tumor removed surgically is overwhelmingly likely to survive. A patient who receives the same risk flag and takes no action will, in a meaningful fraction of cases, present with advanced-stage disease and face a prognosis measured in months.

This arithmetic makes the last mile not merely a logistical problem but an ethical one. An early detection program that generates risk flags it cannot convert into diagnostic resolution is performing an incomplete service. REDI must devote equal effort to the behavioral and structural infrastructure needed to close the loop.

What the Evidence Says About Improving Adherence

Four decades of adherence research have produced a substantial body of evidence about what works. The core finding, replicated across conditions and settings, is that single-component interventions are modestly effective and multi-component interventions are substantially more effective.

Patient Navigation: The Most Consistently Effective Intervention

Patient navigation — the deployment of trained individuals to guide patients through the barriers between a risk signal and completed care — is the most thoroughly studied and consistently effective adherence intervention in oncology. A 2024 systematic review of 59 studies found that 70 percent of studies examining treatment initiation reported significant improvement among navigated patients, 71 percent of studies focused on treatment adherence showed significant gains, and 87 percent of patient satisfaction studies reported meaningful benefit.¹⁵

Navigation works by addressing barriers in real time. Navigated patients receive appointment scheduling assistance, transportation support, insurance navigation, translation services, and emotional encouragement. In breast cancer, patients receiving culturally tailored navigation had shorter times from diagnosis to treatment initiation compared to controls — a difference of 17 to 26 days in predominantly low-income populations.¹⁶

One randomized trial used predictive modeling to identify patients at highest risk of missing appointments and directed navigators specifically to that group. The no-show rate in the intervention group fell from 17.5 percent to 10.2 percent — a 42 percent relative reduction — and navigator calls that reached a patient directly were associated with the most dramatic improvements.¹⁷

Multicomponent and Structural Interventions

The systematic review literature is unambiguous: multicomponent interventions consistently outperform single-component ones. A meta-analysis of 73 randomized trials found that mailed outreach combined with patient navigation improved adherence to repeated stool-based colorectal cancer screening by a larger margin than any single-component intervention.¹⁸ Structural interventions — eliminating copays, offering home-based screening options, and providing single-pill combination therapies — are among the most effective at reaching patients who face cost and access barriers.

Pharmacist-Led Interventions

Pharmacist involvement in outpatient care has been consistently shown to improve adherence and clinical outcomes across chronic disease categories. Telepharmacy and value-based pharmacy systems that allow pharmacists to monitor and document adherence longitudinally represent a scalable infrastructure layer for chronic disease management, particularly for complex medication regimens.

AI-Powered Communication: The Frontier of Adherence Research

The emergence of large language models and conversational AI represents the most significant shift in adherence intervention methodology in a generation. AI-powered tools can communicate in any language, at any hour, at scale across millions of patients simultaneously, without the staffing constraints that limit traditional patient navigation programs. Research in this area is still young — but the pace of published trials has accelerated sharply, and the early results are striking.

What Is Currently Being Studied

The most rigorous completed trial of AI-powered communication in cancer care is the BRIDGE randomized clinical trial (Broadening the Reach, Impact, and Delivery of Genetic Services), conducted at the University of Utah Health and NYU Langone Health between August 2020 and August 2023. The trial enrolled 3,073 patients eligible for cancer genetic evaluation and randomized them to a chatbot-based pretest genetics education program or to standard-of-care genetic counseling appointments.¹⁹

The result was a milestone in the field: the chatbot was statistically equivalent to standard-of-care genetic counseling in completion of pretest genetic services (a difference of just 2.0 percentage points, 95% CI −1.1 to 5.0) and in completion of genetic testing (difference of −1.3 percentage points, 95% CI −3.7 to 1.1). This was not merely a satisfactory outcome — it was a proof of concept that a scalable automated system can deliver cancer-related risk communication and service uptake at the same level as a trained human counselor.¹⁹

Crucially, a secondary analysis of the BRIDGE trial found that social vulnerability — a composite measure of income, housing, employment, and education — did not significantly moderate the chatbot’s effectiveness, suggesting that the tool performed comparably across socioeconomic strata.²⁰ The BRIDGE research team has since received additional funding from the National Cancer Institute to study the chatbot specifically in Spanish-speaking women using mobile cancer screening clinics, bringing the tool to patients rather than requiring patients to come to a clinic.

A separate 2025 retrospective study from WellSpan Health in Pennsylvania and Maryland examined an AI-powered multilingual phone outreach program for colorectal cancer screening among 1,878 patients, including 517 Spanish-speaking patients who lacked active web-based health profiles. The results challenged conventional assumptions: Spanish-speaking patients demonstrated significantly higher engagement than English-speaking patients across all measures — a 2.6-fold higher FIT test opt-in rate (18.2% vs. 7.1%, p<.001), higher connect rates (69.6% vs. 53.0%, p<.001), and longer call duration. The investigators concluded that language-concordant AI interactions may help address longstanding disparities in preventive care access.²¹

In the medication adherence domain, a clinical trial of a voice-based conversational AI application for type 2 diabetes patients found that insulin adherence rates were 32.7 percent higher in the AI group compared to standard care.²² AI-driven SMS reminder systems have also demonstrated meaningful improvements: a study using an AI-generated SMS platform showed significantly higher medication refill rates in older patients with non-communicable diseases compared to control groups that received no reminders.²²

In oncology patient education, studies report that patients interacting with AI chatbots value the immediacy of responses, the ability to ask the same question multiple times without embarrassment, and the non-judgmental tone. Participants in these studies demonstrated improved recall of information about side effects and were more likely to adhere to recommended screening intervals.²³

Who Is Leading the Research

The field is being advanced by investigators across academia, government, and industry, often in collaboration with major health systems:

University of Utah Huntsman Cancer Institute and NYU Langone Health: Kimberly Kaphingst and Guilherme Del Fiol lead the BRIDGE program, the most rigorous existing trial of chatbot-delivered cancer genetic services. Their follow-on NCI-funded work focuses on Spanish-speaking populations.

WellSpan Health (Pennsylvania and Maryland): Published in 2025, this system’s multilingual AI outreach study demonstrated that AI phone agents can achieve higher engagement among Spanish-speaking patients than traditional outreach methods, reversing expected disparity patterns.

National Cancer Institute: Actively funding multiple trials on AI-assisted cancer screening follow-up, including the 2023 JAMA trial on EHR-plus-navigator outreach after abnormal screening results.

American Society of Clinical Oncology (ASCO): Published six guiding principles for AI in oncology in May 2024, establishing an ethical framework for AI tools used in cancer care communication.

JMIR Publications: The leading peer-reviewed journal ecosystem for digital health research, publishing the majority of trials on AI adherence tools in oncology and chronic disease management.

What the Research Shows About Limitations

AI-based adherence tools are not without important caveats. Chatbots that lack robust transparency mechanisms risk allowing patients to believe they are conversing with a human counselor, which raises ethical concerns and can undermine trust when the deception is discovered.²³ AI systems trained on non-diverse datasets may perform worse for populations not represented in their training data, potentially widening disparities rather than narrowing them.²⁴

The “digital divide” remains real. While smartphone ownership exceeds 95 percent among most U.S. demographic groups, health-app engagement is far lower — particularly among older adults, rural populations, and those with limited digital literacy.²⁵ Among low-income immigrant Latino families, studies show that nearly all parents own smartphones and that frequent text messaging is common, but fewer than 10 percent frequently use health-oriented applications.²⁵ AI tools designed without these patterns in mind will fail the patients who most need them.

A 2025 review in The Lancet European Region highlighted specific concerns about undocumented immigrants, who may avoid engagement with any digital health tool that requires identifying information, due to fear of data disclosure to immigration authorities.²⁶ For REDI, which will serve diverse urban populations with large immigrant communities, these privacy and trust considerations are not peripheral — they are central design constraints.

The Direction of the Field

The trajectory of AI in adherence is toward greater personalization, linguistic coverage, and integration with EHR workflows. Natural language processing tools with real-time translation capabilities have shown communication error reductions of 43 percent during clinical encounters, along with 28 percent improvements in patient satisfaction among non-English-speaking populations.²⁷

The next generation of AI adherence tools will likely combine several capabilities that are currently siloed: predictive risk scoring (to identify which patients are most likely to drop out of follow-up), personalized multilingual outreach (tailored to the patient’s language, literacy level, and cultural context), real-time EHR integration (to trigger outreach automatically when a high-risk flag is generated), and human escalation protocols (to route the highest-risk patients to a navigator when AI outreach does not produce action).

This is precisely the architecture that REDI needs. The research base to support it exists and is growing. The task now is to build the implementation infrastructure — and to contribute to the research base in ways that specifically address the REDI population: diverse, urban, multilingual, with high rates of uninsurance and socioeconomic disadvantage.

The REDI Adherence Challenge: A Framework for the Last Mile

REDI’s early detection program introduces an adherence challenge that standard medication-adherence literature does not fully address. When a physician tells a patient with hypertension to take a pill, the patient has experienced symptoms and can observe the consequences of non-adherence over time. When REDI’s algorithm flags a patient as high-risk for lung cancer based on a CBC pattern, the patient typically feels healthy. The urgency of action is invisible.

This is the asymptomatic adherence problem, and it is particularly acute in early detection. The patient must trust a statistical signal enough to undergo a diagnostic workup they would not otherwise have sought. Communicating that trust — translating a model score into a meaningful, actionable message — is as much a behavioral science challenge as a clinical one.

Risk Communication and Framing

The way a high-risk notification is communicated substantially affects whether the patient acts on it. Risk presented as a probability is typically less motivating than the same risk communicated concretely. Physicians and health systems deploying REDI’s algorithms must invest in evidence-based risk communication protocols as a core design requirement. The language used in EHR alerts, patient portal notifications, and physician scripts shapes adherence outcomes before the patient ever reaches the specialist’s office.

Accountability and Tracking Infrastructure

For REDI’s last-mile problem to be solved at scale, health systems must build the infrastructure to track patients from initial risk flag to diagnostic resolution. The INSIGHT Clinical Research Network — spanning NewYork-Presbyterian, Columbia, Montefiore, Mount Sinai, NYU Langone, Houston Methodist, and Stony Brook Medicine — represents the kind of integrated data infrastructure that could, in principle, track patients through the entire follow-up journey. Building last-mile tracking into the REDI deployment architecture from the outset, rather than retrofitting it later, is the correct sequencing.

Establishing the REDI Last Mile Division: A Proposed Plan

REDI’s existing work focuses on algorithm development and validation — the science of producing accurate, early risk signals from routine blood panels. That is essential work, and it must continue. But the survival benefit of a risk signal is realized only when the patient who receives it completes the journey to diagnosis, treatment, and surveillance. No investment in algorithm accuracy matters if patients do not act on the output.

REDI should establish a dedicated organizational unit — the REDI Last Mile Division — with a specific mandate: to ensure that every patient flagged as high-risk by a REDI algorithm completes the care journey on time. The division would combine patient navigation, AI-powered communication, behavioral science research, and policy advocacy into a single operational and research program.

Mission and Structure

The REDI Last Mile Division would operate with four functional components:

AI Communication Lab: Development, validation, and deployment of AI-powered multilingual outreach tools tailored to REDI’s patient population.

Navigation and Care Coordination Team: A trained navigator workforce that handles cases AI outreach cannot resolve, focused on the highest-risk patients and underserved populations.

Behavioral and Social Science Research Unit: A research program studying adherence to asymptomatic algorithmic risk notifications — a largely unstudied problem — with particular attention to urban, immigrant, and low-income populations.

Policy and Reimbursement Advocacy Team: Work with payers, CMS, and state Medicaid programs to establish reimbursable billing codes for AI-assisted outreach and patient navigation in algorithmic early detection programs.

The Comprehensive Operational Plan for Urban, Diverse Populations

The populations served by REDI’s primary institutional partners — NewYork-Presbyterian, Montefiore, NYU Langone, Mount Sinai, and the broader INSIGHT network — are among the most diverse in the United States. Many patients speak a primary language other than English. Many are uninsured or underinsured. Significant proportions are recent immigrants, some undocumented, who carry cultural mistrust of healthcare institutions rooted in historical experience. Many are working multiple jobs, have unreliable transportation, and cannot take time off work for medical appointments. Low health literacy is common. These are not peripheral considerations. They define the problem.

The REDI Last Mile Division’s operational plan must be built around the realities of this population, not the idealized patient of a clinical trial. The following components are required:

Component 1: AI-Powered Multilingual Outreach System

REDI should develop or license an AI communication platform capable of conducting patient outreach in at least 12 languages spoken in its target service areas, including Spanish, Mandarin, Cantonese, Bengali, Haitian Creole, Russian, Arabic, Polish, Tagalog, Korean, Yiddish, and Fuzhounese. The WellSpan multilingual AI study demonstrated that Spanish-speaking patients respond to AI phone outreach at 2.6 times the rate of English-speaking patients who receive comparable outreach through standard channels.²¹ Language-concordant communication is not a courtesy — it is a clinical intervention.

The outreach system should use graduated contact: a patient portal message first, followed by an AI-powered phone call in the patient’s preferred language, followed by a text message with a direct scheduling link, followed by escalation to a human navigator if no response is received within 14 days. Contact timing should be calibrated to patient work schedules, with evening and weekend outreach available for patients who cannot respond during business hours.

All outreach content should be written at a sixth-grade reading level, use concrete language rather than statistical probabilities, and include culturally adapted framing — not a direct translation of English-language text but a culturally appropriate message developed with community input. For populations with lower smartphone ownership or digital engagement, voice-based outreach should be the primary channel, not the fallback.

Component 2: Tiered Patient Navigation

The AI outreach system will not reach every patient. For patients who do not respond to automated outreach after 21 days, or who respond but express confusion, fear, or barriers to care, a trained patient navigator should be assigned. Navigators should be:

Bilingual or multilingual, matching the language composition of the patient population in each health system

Drawn from the communities they serve, with shared cultural background when possible, to build trust more rapidly

Trained specifically in the REDI risk notification framework and in motivational interviewing techniques adapted for asymptomatic risk communication

Empowered to address social determinants of health directly: connecting patients to transportation assistance, childcare support, insurance enrollment, and appointment scheduling on the patient’s behalf

Supported by predictive models that identify which patients are at highest risk of dropout, so navigator resources are directed where they will have the greatest impact

For undocumented patients or patients with immigration-related fears, navigators must be trained in privacy protocols and must be able to reassure patients — credibly and accurately — that their health information is protected under HIPAA and will not be shared with immigration authorities. Community health workers embedded in immigrant communities and trusted by those communities should be a component of the navigation workforce.

Component 3: Asymptomatic Adherence Research Program

REDI should establish a formal research program studying a phenomenon that is almost entirely absent from the adherence literature: patient response to an asymptomatic algorithmic risk notification. The specific questions this research program should address include:

What proportion of patients who receive a REDI high-risk notification complete the full diagnostic journey within 90 days, stratified by disease target, demographic group, language, and health system?

What emotional and psychological responses does an algorithmic risk notification produce — and do responses differ by cultural background, health literacy level, and prior healthcare experience?

What communication modality (phone, text, patient portal, in-person) and what framing produce the highest rates of follow-through, stratified by demographic group?

What are the optimal timing and frequency of follow-up contact to prevent dropout without producing patient fatigue or avoidance?

What are the long-term effects on patient trust and future screening participation of a high-risk flag that does not result in a cancer diagnosis?

This research program should publish its findings in peer-reviewed journals and contribute to the NCI-funded adherence literature. REDI has a unique opportunity to generate the evidence base that will guide early detection programs globally, because no other program will generate risk signals at comparable scale across a population as diverse as REDI’s.

Component 4: Partnership and Research Collaboration Strategy

The REDI Last Mile Division should pursue active research partnerships with institutions and investigators who are already advancing the field. Specifically:

Huntsman Cancer Institute / University of Utah: Partner with Kimberly Kaphingst and the BRIDGE research team to adapt and deploy their chatbot infrastructure for REDI’s risk notification context, and contribute REDI’s diverse urban patient population to their planned studies of Spanish-speaking populations.

NYU Langone Health: NYU is already a BRIDGE trial site and an INSIGHT network member. A formal collaboration between REDI and NYU’s cancer genetics and health equity research teams would allow rapid deployment of tested AI tools within an existing institutional relationship.

Montefiore Health System and the Albert Einstein College of Medicine: Montefiore is a national leader in serving urban, immigrant, and low-income populations in the Bronx — REDI’s target patient population in concentrated form. Its health equity and community health worker research infrastructure should be a core partner for the Last Mile Division’s urban adherence program.

National Cancer Institute: Apply for NCI R01 or P01 funding specifically for the asymptomatic adherence research program. NCI’s Cancer Moonshot Initiative and its health equity research portfolio represent appropriate funding mechanisms.

Patient-Centered Outcomes Research Institute (PCORI): PCORI funds patient-centered comparative effectiveness research with explicit attention to underserved populations. The REDI Last Mile Division’s research agenda is precisely aligned with PCORI’s priorities.

Community-Based Organizations: Partner with community health centers, federally qualified health centers (FQHCs), and community-based organizations embedded in the immigrant and low-income communities that REDI serves. Community trust is not generated by a health system — it is borrowed from trusted community institutions. The Last Mile Division must be visible and accountable in these communities.

Component 5: Health Literacy and Plain-Language Risk Communication Standards

REDI should develop and publish a standardized risk communication protocol for all high-risk notifications. This protocol should:

Use simple, direct language at a sixth-grade reading level for all patient-facing materials

Avoid statistical framing in favor of concrete language (e.g., “This result means you should see a doctor for one additional test within the next 30 days” rather than “Your probability of early-stage disease is elevated”)

Be reviewed and approved by community advisory boards representing the major demographic groups in the REDI service population

Include culturally adapted versions — not translations but true adaptations — in all 12 major languages of the REDI service area

Be piloted and validated with real patients before deployment, with particular attention to whether patients in low-literacy, non-English-speaking, and high-anxiety populations correctly understand the message and its urgency

Component 6: Social Determinants of Health Integration

Many of the patients REDI will flag as high-risk will face barriers to care that have nothing to do with their willingness to act on a risk notification. They may lack transportation to a specialist appointment. They may be unable to take a day off work for a colonoscopy prep. They may be afraid of the cost of diagnostic testing even with insurance. They may not have insurance at all. The REDI Last Mile Division must address these barriers directly, not treat them as someone else’s problem.

Establish a social needs screening protocol for all high-risk-flagged patients, using validated instruments such as the PRAPARE tool or the AHC Health-Related Social Needs screening tool

Create a resource navigation function that connects patients to transportation assistance (Lyft Health, Uber Health, transit subsidies), appointment translation services, cost-assistance programs, and insurance enrollment support

Partner with health systems to waive copays and cost-sharing for follow-up diagnostic testing triggered by a REDI high-risk notification, making the case that the cost of a colonoscopy is trivial compared to the cost of treating Stage IV colon cancer

For patients with food insecurity, housing instability, or other immediate subsistence needs, connect those needs to community resources as part of the navigation encounter — recognizing that a patient who is worried about rent cannot prioritize a cancer screening appointment

Component 7: Technology Infrastructure

Build or license a centralized last-mile tracking dashboard integrated into Epic and other EHR platforms used by INSIGHT network members, displaying real-time status for every high-risk-flagged patient from notification to diagnostic resolution

Implement automated EHR triggers that initiate AI outreach within 48 hours of a REDI high-risk flag being generated

Use predictive models — built from INSIGHT’s 23 million de-identified EHRs — to stratify patients by dropout risk and allocate navigator resources proportionally

Establish FHIR-compliant API connections that allow the last-mile tracking system to communicate with patients through patient portals (MyChart), SMS, and AI phone agents within a single unified workflow

Ensure that all technology infrastructure meets HIPAA requirements and that patient data is never shared with third parties for commercial purposes — a non-negotiable trust-building requirement for immigrant and underserved populations

Performance Metrics and Accountability

The REDI Last Mile Division should be held accountable to measurable performance standards, reported publicly on an annual basis:

Flag-to-diagnostic-resolution rate: the proportion of high-risk-flagged patients who complete confirmatory diagnostic testing within 90 days of the initial flag

Flag-to-first-contact rate: the proportion of patients who receive and acknowledge a risk notification within 14 days

Equity metrics: flag-to-diagnostic-resolution rates stratified by language, race, ethnicity, insurance status, and health system, with explicit targets to close disparity gaps year-over-year

Treatment initiation rate among confirmed-positive patients: the proportion of patients with a confirmed early-stage cancer who begin treatment within 30 days of diagnosis

Patient-reported experience: standardized survey data on patient experience with the risk notification, follow-up outreach, and navigation process, in all 12 languages

Policy and Reimbursement Advocacy

The REDI Last Mile Division must engage with payers, CMS, and federal and state policy stakeholders to address two critical gaps simultaneously: the absence of CPT reimbursement codes for algorithmic screening analysis (which limits REDI’s commercial sustainability) and the absence of reimbursement for patient navigation in algorithmic early detection programs (which limits the Last Mile Division’s operational sustainability).

The division should work with the American Cancer Society, ASCO, and the National Coalition for Cancer Survivorship to build the policy case for navigator reimbursement. The case is straightforward: the cost of a navigator who ensures that a Stage I colon cancer patient receives timely treatment is a fraction of the cost of treating a Stage IV colon cancer — the outcome that results when the same patient drops out of follow-up.

Conclusion: A Complete System, Not a Detection System

The central insight of the Ratner Early Detection Initiative is that the blood tests drawn at routine annual physicals contain diagnostic intelligence that medicine has not yet learned to read systematically. REDI is building the machine learning infrastructure to read them. But reading the signal is only the first step in a longer chain. The survival benefit of early detection is realized only when the patient who is flagged as high-risk completes the journey from notification to diagnostic resolution to definitive treatment.

The evidence reviewed in this paper is sobering. In the best-studied early detection programs in medicine, real-world follow-up rates after abnormal results fall well short of clinical trial rates. In standard care, fewer than one quarter of patients overdue for follow-up after abnormal cancer screening complete that follow-up within four months.¹⁰ Half of patients on long-term chronic disease therapy do not take their medications as prescribed.¹

These numbers are arguments for building the last-mile infrastructure with the same rigor and ambition that goes into the algorithms themselves. They are also arguments for urgency. The research base now exists to support a serious, well-designed, AI-augmented last-mile program. A multilingual AI phone agent has demonstrated that Spanish-speaking patients engage at rates that exceed English-speaking patients under standard outreach.²¹ A chatbot-delivered cancer genetics service has proven equivalent to human genetic counseling across 3,073 patients in a rigorous randomized trial.¹⁹ Patient navigation reduces no-show rates by 42 percent when targeted to high-risk patients.¹⁷

The REDI Last Mile Division — combining AI-powered multilingual outreach, tiered human navigation, behavioral science research, social determinants screening, community partnerships, and policy advocacy — is the organizational structure that can convert REDI’s detection capability into population-wide survival benefit. Building it is not an optional enhancement. For a program whose purpose is to save lives, it is the mission.

Endnotes

1.Brown MT, Bussell JK. Medication Adherence: WHO Cares? Mayo Clin Proc. 2011;86(4):304-314. [https://doi.org/10.4065/mcp.2010.0575](https://doi.org/10.4065/mcp.2010.057

Widely cited synthesis establishing that approximately 50 percent of patients with chronic conditions do not take their medications as prescribed. Establishes the foundational prevalence statistic underpinning the global scale of non-adherence.

2.McDonnell PJ, Jacobs MR. Hospital Admissions Resulting from Preventable Adverse Drug Reactions. Ann Pharmacother. 2002;36(9):1331-1336. [https://doi.org/10.1345/aph.1A333](https://doi.org/10.1345/aph.1A33

Documents that medication non-adherence contributes to approximately 125,000 preventable deaths annually and up to 25 percent of hospital admissions in the United States, establishing the scale of harm that adherence interventions must address.

3.Kardas P, Bennett B, Borah F, et al. Medication Non-Adherence: Reflecting on Two Decades Since the WHO Adherence Report. Front Pharmacol. 2024;15:1444012. [https://doi.org/10.3389/fphar.2024.1444012](https://doi.org/10.3389/fphar.2024.144401

Twenty-year review of the global medication non-adherence landscape, including quantitative modeling showing that a 20 percent improvement in cardiovascular adherence correlates with an 8 percent reduction in cardiovascular events and 12 percent decrease in mortality.

4.Boye KS, et al. Medication Adherence Star Ratings Measures, Health Care Resource Utilization, and Cost. Am J Manag Care. 2024;30(5):210-217. [https://doi.org/10.37765/ajmc.2024.89528](https://doi.org/10.37765/ajmc.2024.8952

Analysis of Medicare Advantage data demonstrating that patients missing adherence benchmarks for diabetes, hypertension, and hyperlipidemia face 21 to 50 percent higher inpatient stay rates and 11 to 13 percent higher total healthcare costs. Directly supports the economic argument for adherence investment.

5.Aljofan M, Oshibayeva A, Moldaliyev I, et al. The Rate of Medication Nonadherence and Influencing Factors: A Systematic Review. Electron J Gen Med. 2023;20(3):em471. [https://doi.org/10.29333/ejgm/12940](https://doi.org/10.29333/ejgm/1294

Systematic review identifying that six of the eight most common predictors of medication non-adherence are rooted in patient psychology and health literacy, rather than cost or logistics. Supports the behavioral science framing of the adherence challenge.

6.Osterberg L, Blaschke T. Adherence to Medication. N Engl J Med. 2005;353(5):487-497. [https://doi.org/10.1056/NEJMra050100](https://doi.org/10.1056/NEJMra05010

Authoritative NEJM review of the multidimensional determinants of medication adherence, establishing that inadequate explanation of treatment rationale at the point of prescribing is a primary and correctable driver of non-adherence.

7.Rolnick SJ, Pawloski PA, Hedblom BD, et al. Patient Characteristics Associated with Medication Adherence. Clin Med Res. 2013;11(2):54-65. [https://doi.org/10.3121/cmr.2013.1113](https://doi.org/10.3121/cmr.2013.111

Large retrospective study establishing therapy-related factors — including adverse drug reactions, dosing frequency, and side effect burden — as major independent contributors to medication non-adherence across multiple chronic disease categories.

8.Fischer MA, Stedman MR, Lii J, et al. Primary Medication Non-Adherence: Analysis of 195,930 Electronic Prescriptions. J Gen Intern Med. 2010;25(4):284-290. [https://doi.org/10.1007/s11606-010-1253-9](https://doi.org/10.1007/s11606-010-1253-

Large-scale electronic prescription analysis demonstrating the dramatic dose-response relationship between out-of-pocket cost and prescription abandonment — below 5 percent at zero cost, rising to 60 percent when copays exceed $500.

9.Rohatgi KW, Humble S, Hollar L, et al. Medication Adherence and Characteristics of Patients Who Spend Less on Basic Needs to Afford Medications. J Am Board Fam Med. 2021;34(3):561-570. [https://doi.org/10.3122/jabfm.2021.03.200361](https://doi.org/10.3122/jabfm.2021.03.20036

Survey of low-income adults documenting the trade-off between medication cost and basic subsistence needs, identifying patients who sacrifice food or housing to pay for medications as a high-risk group for non-adherence with implications for health system design.

10.Atlas SJ, Tosteson TD, Lowenstein LM, et al. Effect of Reminders and Outreach on Follow-Up of Abnormal Cancer Screening Results. JAMA. 2023;330(14):1357-1366. [https://doi.org/10.1001/jama.2023.17594](https://doi.org/10.1001/jama.2023.1759

NCI-funded randomized trial of nearly 12,000 patients at 44 practices demonstrating that fewer than 25 percent of patients overdue for follow-up after abnormal cancer screening completed testing under usual care — the most direct evidence for the scale of the last-mile failure.

11.Robertson DJ, Lee JK, Boland CR, et al. Recommendations on Fecal Immunochemical Testing to Screen for Colorectal Neoplasia: A Consensus Statement. Gastroenterology. 2017;152(5):1217-1237. [https://doi.org/10.1053/j.gastro.2016.08.053](https://doi.org/10.1053/j.gastro.2016.08.05

Consensus statement establishing that real-world colonoscopy completion rates after a positive fecal immunochemical test fall far short of clinical trial rates, with safety-net data showing only 18 to 56 percent follow-through within one year.

12.Zorzi M, Battagello J, Selby K, et al. Non-Compliance with Colonoscopy after a Positive Faecal Immunochemical Test Doubles the Risk of Dying from Colorectal Cancer. Gut. 2022;71(3):561-567. [https://doi.org/10.1136/gutjnl-2020-323688](https://doi.org/10.1136/gutjnl-2020-32368

Population-based study demonstrating that failure to complete follow-up colonoscopy after a positive stool test doubles the risk of colorectal cancer death — the most direct evidence for the lethal cost of last-mile failure in cancer screening.

13.Doubeni CA, Corley DA, Quinn VP, et al. Expanded Lung and Colorectal Cancer Screening — Ensuring Equity and Safety Under New Guidelines. N Engl J Med. 2021;385(24):2237-2242. [https://doi.org/10.1056/NEJMp2113332](https://doi.org/10.1056/NEJMp211333

NEJM perspective identifying structural gaps in cancer screening follow-up, including the absence of mandatory mechanisms to ensure patients act on abnormal results and the lack of reimbursement for patient navigators in preventive care programs.

14.Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer Statistics, 2023. CA Cancer J Clin. 2023;73(1):17-48. [https://doi.org/10.3322/caac.21763](https://doi.org/10.3322/caac.2176

Annual American Cancer Society statistics documenting stage-specific survival rates for all major cancers, including the stark survival differential between localized (90%) and metastatic (13%) colorectal cancer. Provides the quantitative stakes of the last-mile problem.

15.Battaglia TA, Bak SM, Heeren T, et al. Patient Navigation in Cancer Treatment: A Systematic Review. Curr Oncol Rep. 2024;26(5):452-464. [https://doi.org/10.1007/s11912-024-01514-9](https://doi.org/10.1007/s11912-024-01514-

Systematic review of 59 studies finding that 70 percent showed improved cancer treatment initiation, 71 percent improved adherence, and 87 percent improved patient satisfaction among navigated patients. The strongest available evidence base for patient navigation as a last-mile intervention.

16.Chan RJ, Johal H, Crawford-Williams F, et al. Patient Navigation Across the Cancer Care Continuum: An Overview of Systematic Reviews. CA Cancer J Clin. 2023;73(6):561-589. [https://doi.org/10.3322/caac.21788](https://doi.org/10.3322/caac.2178

Umbrella review of 26 systematic reviews synthesizing evidence from 172 primary studies, finding strong evidence that patient navigation improves cancer screening rates and reduces time from diagnosis to treatment initiation, including in diverse and disadvantaged populations.

17.Lasser KE, Murillo J, Medlin E, et al. Patient Navigation Based on Predictive Modeling Decreases No-Show Rates in Cancer Care. Cancer. 2015;121(10):1662-1670. [https://doi.org/10.1002/cncr.29236](https://doi.org/10.1002/cncr.2923

Randomized trial demonstrating that telephone navigation targeted at patients predicted to be at highest no-show risk reduced missed appointment rates from 17.5 percent to 10.2 percent — a 42 percent relative reduction — using a predictive model to allocate navigator resources efficiently.

18.Holden DJ, Jonas DE, Porterfield DS, et al. Systematic Review: Enhancing the Use and Quality of Colorectal Cancer Screening. Ann Intern Med. 2010;152(10):668-676. [https://doi.org/10.7326/0003-4819-152-10-201005180-00239](https://doi.org/10.7326/0003-4819-152-10-201005180-0023

Meta-analysis of 73 randomized trials establishing that multicomponent interventions — combining mailed outreach, patient reminders, and navigation — consistently outperform single-component approaches in improving cancer screening adherence. Establishes the evidence base for bundled last-mile interventions.

19.Kaphingst KA, Kohlmann WK, Lorenz Chambers R, et al. Uptake of Cancer Genetic Services for Chatbot vs Standard-of-Care Delivery Models: The BRIDGE Randomized Clinical Trial. JAMA Netw Open. 2024;7(9):e2432143. [https://doi.org/10.1001/jamanetworkopen.2024.32143](https://doi.org/10.1001/jamanetworkopen.2024.3214

Landmark equivalence trial of 3,073 patients demonstrating that a chatbot-based cancer genetics education program was statistically equivalent to standard-of-care human genetic counseling for completion of pretest services and genetic testing — a proof of concept for scalable AI-delivered cancer risk communication.

20.Bather JR, Goodman MS, Harris A, et al. Social Vulnerability, Chatbots, and Genetic Service Utilization Among Unaffected Patients with Inherited Cancer Susceptibility in the BRIDGE Trial. BMC Cancer. 2025;25:180. [https://doi.org/10.1186/s12885-025-13495-4](https://doi.org/10.1186/s12885-025-13495-

Secondary analysis of the BRIDGE trial finding that social vulnerability did not significantly moderate chatbot effectiveness, suggesting AI-delivered genetic counseling performs comparably across socioeconomic strata. Critical evidence for deploying AI tools in disadvantaged urban populations.

21.Bhimani M, Baker RH, Ausin MS, et al. Using a Multilingual AI Care Agent to Reduce Disparities in Colorectal Cancer Screening. J Med Internet Res. 2025;27:e71211. [https://doi.org/10.2196/71211](https://doi.org/10.2196/7121

Retrospective study of 1,878 patients at WellSpan Health demonstrating that Spanish-speaking patients showed 2.6-fold higher FIT test opt-in rates than English-speaking patients when reached by a multilingual AI conversational agent, challenging assumptions about digital tool effectiveness in non-English-speaking populations.

22.Digital Interventions in Medication Adherence: A Narrative Review of Current Evidence and Challenges. PMC. 2024. [https://pmc.ncbi.nlm.nih.gov/articles/PMC12549263/](https://pmc.ncbi.nlm.nih.gov/articles/PMC12549263

Narrative review synthesizing evidence that AI voice applications improved insulin adherence by 32.7 percent over standard care, and that AI-generated SMS reminders significantly improved medication refill rates in older patients with chronic conditions — establishing the effectiveness hierarchy of digital adherence tools.

23.Thind BS, Tsao CK. Artificial Intelligence in Oncology: Promise, Peril, and the Future of Patient-Physician Interaction. Front Digit Health. 2025;7:1633577. [https://doi.org/10.3389/fdgth.2025.1633577](https://doi.org/10.3389/fdgth.2025.163357

Comprehensive review of AI tools in oncology care, documenting that patients value AI chatbots for immediacy, non-judgmental tone, and ability to ask questions repeatedly. Also documents transparency and misinformation risks that require design safeguards in any patient-facing AI adherence tool.

24.Frontiers. A Critical Look into Artificial Intelligence and Healthcare Disparities. Front Artif Intell. 2025. [https://doi.org/10.3389/frai.2025.1545869](https://doi.org/10.3389/frai.2025.154586

Critical analysis of AI in healthcare disparities, documenting that AI systems trained on non-diverse datasets may perform worse for underrepresented populations — a specific risk for tools deployed in diverse urban environments without careful attention to training data representation.

25.López-Cevallos D, et al. Information and Communication Technology Access and Use Among Low-Income Latino Immigrant Parents. Matern Child Health J. 2021. [https://doi.org/10.1007/s10995-021-03265-6](https://doi.org/10.1007/s10995-021-03265-

Survey study finding that nearly all low-income Latino immigrant parents own smartphones and use text messaging frequently, but fewer than 10 percent use health-oriented applications — establishing the gap between device ownership and health-app engagement that AI adherence tools must bridge.

26.Takian A, Zaman MH, et al. Artificial Intelligence in Migrant Health: A Critical Perspective on Opportunities and Risks. Lancet Reg Health Eur. 2025;57:101421. [https://doi.org/10.1016/j.lanepe.2025.101421](https://doi.org/10.1016/j.lanepe.2025.10142

Critical review documenting that undocumented immigrants avoid digital health tools that require identifying information due to fear of data disclosure to immigration authorities — a central design constraint for AI adherence programs serving diverse urban populations including immigrant communities.

27.Bridging the Digital Divide: Artificial Intelligence as a Catalyst for Health Equity in Primary Care Settings. ScienceDirect. 2025. [https://doi.org/10.1016/j.ijmedinf.2025.001](https://doi.org/10.1016/j.ijmedinf.2025.00

Narrative review finding that AI-based translation tools reduced communication errors by 43 percent during clinical encounters and improved patient satisfaction by 28 percent among non-English-speaking populations — demonstrating the clinical value of linguistically appropriate AI communication tools.