Using routine clinical features to classify adult-onset diabetes at diagnosis: the StartRight prospective observational study

The StartRight prospective observational study demonstrates that a classification model combining routine clinical features, such as lower age and BMI, unintentional weight loss, and high presentation glycaemia, with or without islet autoantibodies, achieves high accuracy in differentiating adult-onset type 1 from type 2 diabetes at diagnosis, outperforming current clinical guidance.

The Gist

The Big Problem: A Case of Mistaken Identity

Imagine walking into a hospital with a fever. The doctor has to decide: Is this the flu, or is it a bacterial infection? The treatment for each is completely different. If you get the wrong medicine, you could get much sicker.

This is exactly the problem with adult-onset diabetes. There are two main types: Type 1 and Type 2.

• Type 2 is like a clogged pipe; the body is resistant to insulin, but it still makes some. It's very common in adults.
• Type 1 is like a broken factory; the body's immune system has destroyed the factory that makes insulin. It is less common in adults but happens more often than people think.

The trouble is, when an adult gets diabetes, doctors often guess which type it is based on age or weight. But this "guessing game" leads to mistakes. About one in three adults who actually have Type 1 are mistakenly told they have Type 2 and treated with the wrong medicine.

The Study: Building a Better Detective Kit

The researchers behind this study (the StartRight team) wanted to stop the guessing. They asked: "What specific clues, available right when a patient walks in, can tell us exactly which type of diabetes they have?"

They didn't just look at one clue; they looked at 11 different routine features (like age, weight, waist size, and blood sugar levels) and combined them into a "detective kit."

The Key Clues (The "Smoking Guns")

The study found that while doctors already knew age and weight mattered, they missed some other huge clues. The most powerful indicators that a patient has Type 1 (the "broken factory") were:

1. Younger age (at diagnosis).
2. Lower Body Mass Index (BMI) (being thinner).
3. Lower waist-to-hip ratio (carrying less fat around the middle).
4. Unintentional weight loss (dropping pounds without trying).
5. Very high blood sugar at the moment of diagnosis.

The Metaphor: Think of diagnosing diabetes like trying to identify a car by its engine noise.

• Old Method: "It sounds like a truck, so it must be a truck." (Based only on age/weight).
• New Method: "It sounds like a truck, but it's missing a key part, it's losing fuel rapidly, and it's vibrating differently." (Combining all the clues).

The Solution: The "StartRight Score"

The researchers didn't just list the clues; they built a calculator (called the StartRight Score).

• How it works: You plug in the patient's routine numbers (age, weight, blood sugar, etc.).
• The Result: The calculator gives a score.
  • Low Score: "This looks like Type 2."
  • High Score: "This looks like Type 1."
  • Middle Score: "We aren't sure; let's run a specific lab test (antibody test) to be safe."

Why This is a Game-Changer

The study tested this calculator in two ways:

1. In a controlled study: They followed patients for years to see if the calculator was right. It was extremely accurate (94–97% accuracy), far better than just looking at age or weight alone.
2. In real-world data: They tested it on records from over 180,000 people in the UK. They found that people who the calculator flagged as "likely Type 1" actually went on to need insulin quickly or had other signs of Type 1, even if their doctors had initially treated them as Type 2.

The "Filter" for Lab Tests

Currently, guidelines suggest testing for specific antibodies (the "smoking gun" for Type 1) in many people. However, these tests cost money and take time.

The study suggests using the StartRight Score as a filter:

• If the score says "Very Low Chance of Type 1," you probably don't need the expensive lab test.
• If the score says "High Chance," you should definitely get the test.

This helps doctors prioritize who needs the test, saving money and ensuring the right people get the right diagnosis faster.

The Bottom Line

This study proves that we don't need to guess. By combining simple, routine facts (like how much weight a person lost or their waist size) into a single score, we can tell the difference between Type 1 and Type 2 diabetes in adults with very high accuracy. This helps ensure adults get the correct treatment immediately, rather than being misdiagnosed and treated for the wrong condition.

Technical Summary

1. Problem Statement

The differentiation between Type 1 Diabetes (T1D) and Type 2 Diabetes (T2D) in adults at the time of diagnosis is a significant clinical challenge. While T1D is often associated with younger age and T2D with older age and obesity, these conditions overlap substantially in adults.

• Misclassification Rates: Approximately one in three adults with T1D are initially misdiagnosed and treated as T2D, while one in six adults diagnosed with T1D actually have T2D.
• Limitations of Current Guidance: Previous systematic reviews (2015, 2022) identified only age-at-diagnosis and BMI as robust evidence-based discriminators. Other features commonly cited in guidelines (e.g., osmotic symptoms, family history) lack strong empirical support.
• Biomarker Constraints: While islet-autoantibodies and C-peptide are gold standards for classification, C-peptide has limited utility at diagnosis, and routine islet-autoantibody testing is cost-prohibitive for all new-onset adult diabetes cases. Furthermore, antibody presence can overlap between T1D and T2D, making interpretation dependent on pre-test probability.

Objective: To identify routine clinical features that discriminate between adult-onset T1D and T2D at diagnosis and to develop high-accuracy classification models (with and without biomarkers) to guide clinical decision-making and testing prioritization.

2. Methodology

Study Design:
A prospective observational cohort study (StartRight) with an additional validation cohort (StartRight Prime) and external validation using UK primary care data (CPRD).

Participants:

• StartRight Cohort: 1,800 adults (aged 18–50 years) diagnosed with diabetes within the previous 12 months.
• StartRight Prime Cohort: 470 adults (aged >50 years) diagnosed within the previous 12 months. This cohort was enriched for T1D cases (aiming for a 1:1 insulin-treated vs. non-insulin-treated ratio) to ensure sufficient statistical power for the rarer condition in older adults.
• Exclusion: Gestational, monogenic, or secondary diabetes.

Data Collection:

• Baseline: Routine clinical features were recorded at recruitment (within 12 months of diagnosis), including age, BMI, waist-hip ratio, unintentional weight loss, osmotic symptoms, ethnicity, family history (insulin vs. non-insulin treated), autoimmune history, and presentation glycaemia (HbA1c, glucose, DKA status).
• Biomarkers: Islet-autoantibodies (GAD, IA2, ZnT8) and Type 1 Diabetes Genetic Risk Score (T1DGRS) were measured at recruitment.
• Outcome Definition (Gold Standard): Diabetes subtype was defined ≥3 years post-diagnosis based on:
  • T1D: Continuous insulin treatment within 3 years AND C-peptide <600 pmol/L.
  • T2D: No insulin treatment OR (if on insulin) C-peptide ≥600 pmol/L.
  • Sensitivity Analysis: A restrictive definition incorporating islet-autoantibody status was also used.

Statistical Analysis:

• Feature Selection: Logistic regression was used to assess which routine clinical features added significant discriminatory power beyond age and BMI.
• Model Development: Multivariable logistic regression models were built using:
  1. Clinical features only.
  2. Clinical features + Islet-autoantibodies.
  3. Clinical features + T1DGRS.
• Validation: Internal validation via bootstrapping (1,000 iterations) to assess optimism and stability. External validation performed in the StartRight Prime cohort and the CPRD population dataset.
• Performance Metrics: Area Under the Receiver Operating Characteristic Curve (AUCROC), Precision-Recall (AUCPR), sensitivity, specificity, and calibration curves.
• Scoring System: A simplified points-based score ("StartRight Score") was derived by rounding model beta coefficients.

3. Key Contributions

1. Identification of Discriminative Features: The study identified 11 routine clinical features that significantly differentiate T1D from T2D. The most discriminative features (after age and BMI) were:
   • Unintentional weight loss.
   • Waist-hip ratio.
   • Presentation glycaemia (HbA1c or glucose).
   • Other features like osmotic symptoms and family history of autoimmunity were found to be weakly discriminative.
2. Development of High-Accuracy Prediction Models:
   • Clinical Features Only Model: Achieved an AUCROC of 0.94 (95% CI: 0.93–0.96) in the 18–50 age group and 0.93 in the >50 age group.
   • Clinical + Antibody Model: Improved performance to an AUCROC of 0.97 (95% CI: 0.96–0.98).
   • StartRight Score: A simplified point-based tool maintained high performance (AUCROC 0.94), offering a pragmatic alternative to complex calculators.
3. Optimization of Biomarker Testing: The study demonstrated that using the clinical model to stratify patients allows for the prioritization of islet-autoantibody testing. This approach significantly reduces the number of patients requiring expensive testing while maintaining high sensitivity for detecting T1D compared to current NICE and ADA/EASD guidelines.
4. Population-Level Utility: Validation in the CPRD dataset (over 1.6 million people) confirmed that high model probabilities strongly predicted outcomes associated with misclassified T1D (e.g., rapid progression to insulin, ketoacidosis) even in patients initially treated as T2D.

4. Key Results

• Feature Performance: No single clinical feature adequately differentiated T1D from T2D in isolation (highest single-feature accuracy was 79% for BMI <28 kg/m²). However, combining features yielded superior results.
• Model Discrimination:
  • StartRight (18–50 yrs): Clinical features only AUC = 0.94; Clinical + Antibodies AUC = 0.97.
  • StartRight Prime (>50 yrs): Clinical features only AUC = 0.93; Clinical + Antibodies AUC = 0.97.
  • The models maintained high discrimination across subgroups, including different ethnicities (Black, South Asian) and sexes, though sample sizes for some subgroups were small.
• Comparison with Guidelines:
  • Current guidelines (NICE, ADA/EASD) trigger testing based on single criteria (e.g., age <50, BMI <25).
  • The StartRight Score (cut-off ≥7 points) and Clinical Model (probability ≥10%) identified a similar proportion of T1D cases as "universal testing" but required testing in a much smaller subset of the population, thereby improving specificity and resource efficiency.
• CPRD Validation: In the UK primary care dataset, patients with a high probability of T1D (based on the clinical model) were significantly more likely to develop ketoacidosis, require basal-bolus insulin, or receive a T1D diagnostic code within 3 years, regardless of initial treatment.

5. Significance and Implications

• Clinical Practice: The study provides the first evidence-based, validated tool to assist clinicians in classifying adult-onset diabetes at diagnosis. It moves beyond reliance on age and BMI alone.
• Resource Allocation: By using the StartRight Score or model probability, healthcare systems can prioritize islet-autoantibody testing for patients with a high pre-test probability of T1D. This avoids the cost and logistical burden of testing low-probability patients while ensuring high-risk patients are not missed.
• Guideline Revision: The findings suggest that current guidelines may be too broad (testing too many low-risk patients) or too narrow (missing T1D in older/obese adults). The new models offer a more nuanced, data-driven approach.
• Future Implementation: The authors recommend the integration of these models into electronic health records (EHR) as automated decision support tools to facilitate real-time classification and treatment planning.

Limitations: The study relied on self-reported data for some features (e.g., weight loss), and the >50 age cohort was enriched for T1D, meaning absolute prevalence calibration in the general older population requires further validation. Additionally, the CPRD analysis could not incorporate antibody or genetic data due to data unavailability.

Conclusion:
The StartRight study successfully demonstrates that a combination of routine clinical features—specifically age, BMI, waist-hip ratio, unintentional weight loss, and presentation glycaemia—can accurately classify adult-onset diabetes. These models outperform current clinical guidance and single biomarker testing, offering a scalable solution to reduce misclassification and optimize the use of classification biomarkers.