Identification of Patient Clusters with Distinct Disease Progression Patterns Utilizing a Nationwide Finnish Population with Type 2 Diabetes

Abstract

Introduction

Variability in the clinical presentation of patients with type 2 diabetes (T2D) is high and underlines the need for more personalized patient care. This nationwide study aimed to describe characteristics, treatment patterns, and disease progression of Finnish patients with T2D (N = 302,987), and to identify patient clusters with distinct progression patterns based on the occurrence of diabetes-related complications.

Methods

The study included all adult patients with incident T2D in Finland between 2010 and 2019. Data were collected from national health and social care registers, data lakes, and a private healthcare provider between 1996 and 2021. Patient clusters were identified based on disease progression, defined by the occurrence of 22 pre-defined end-points, using likelihood-based growth mixture modeling.

Results

Five patient clusters with stable (C1; n = 133,951), mild (C2; n = 52,819), moderate (C3, n = 43,488), rapid (C4; n = 10,159), and extremely rapid progression (C5; n = 1973) were identified. The mean number of end-point complications per patient at baseline ranged from 0.2 to 2.3 across clusters and remained stable in C1–C3 over the first 5 years. In C5, the number increased to 5.5 and 7.2 during the first and third follow-up years, respectively, with a similar but more modest annual increase observed in C4. Cardiovascular complications increased more rapidly in C5 and C4 than C1–C3. T2D medication use was more common in milder clusters, whereas 31.4% and 48.2% of patients in C4 and C5, respectively, had no T2D medication. The rate of certain infections and values of creatinine, hemoglobin, and erythrocytes, increased with cluster severity.

Conclusions

Diagnosis of several other new conditions, particularly cardiovascular complications, at or soon after incident T2D diagnosis predicts poor prognosis. The results further support a comprehensive approach in diabetes care, including evaluation and treatment of cardiovascular diseases alongside glycemic control.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13300-026-01879-z.

Plain Language Summary

It is well known that people with type 2 diabetes can experience the disease in different ways, with wide variation in symptoms and disease progression. To support more personalized care, clinically useful tools that can predict how the disease will develop are needed. The researchers analyzed nationwide health register data from Finland, including all patients who received their first diagnosis of type 2 diabetes between 2010 and 2019. Their goal was to group patients into clusters with different long-term disease progression patterns, based on the development of diabetes-related complications after diagnosis. In total, the study included 302,987 patients with type 2 diabetes. These patients were divided into five clusters showing either stable (cluster 1), mild (cluster 2), moderate (cluster 3), rapid (cluster 4), or extremely rapid (cluster 5) disease progression. Further analysis showed that having additional conditions diagnosed at, or shortly after, the onset of diabetes was linked to a worse prognosis. Notably, 18.6% of patients did not purchase any diabetes medication during the follow-up period. This proportion was particularly high in patients allocated to cluster 4 (31.4%) and cluster 5 (48.2%). This study shows that patients with type 2 diabetes can be classified into meaningful groups with different progression patterns. This approach may help clinicians make timely and cost-effective decisions, focusing resources on patients most likely to benefit from early and intensive treatment.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13300-026-01879-z.

Key Summary Points

Why carry out this study?

Identifying factors that predict type 2 diabetes progression at an early stage would aid personalized treatment in diabetes care.

Can patients with type 2 diabetes be allocated into clinically meaningful clusters based on the occurrence of type 2 diabetes-related end-points?

What was learned from the study?

Patients with type 2 diabetes were allocated into five clinically meaningful clusters with stable (cluster 1), mild (cluster 2), moderate (cluster 3), rapid (cluster 4) and extremely rapid (cluster 5) long-term disease progression profiles.

Classifying patients into prognostically meaningful clusters could help clinicians make timely, cost-effective decisions in patient care by allocating resources to those who would benefit most.

Introduction

Diabetes is a rapidly growing global health concern, with over 580 million affected adults in 2024—a number expected to exceed 853 million by 2050 [1]. The most prominent types of diabetes are type 1 diabetes, caused by autoimmune beta-cell destruction leading to full insulin deficiency, and type 2 diabetes (T2D), characterized by beta-cell dysfunction, insulin resistance, and metabolic syndrome and accounting for over 90% of diabetes cases worldwide [1, 2]. In Finland, approximately 500,000 adults live with diabetes, of which 400,000 are diagnosed with T2D [3]. T2D varies in presentation and can lead to serious complications, including cardiovascular disease (CVD) and conditions from capillary damage, such as retinopathy, chronic kidney disease, and neuropathy [1].

T2D and its complications pose a significant economic burden on healthcare systems. In 2007, diabetes accounted for around 15% of Finland’s total healthcare costs, and patients with complications are known to have 2.2 to 12 times higher costs than those without complications [4–6]. Identification of risk factors for the development of T2D enables timely diagnosis and glucose management, whereas identifying specific factors that predict disease progression at an early stage would be crucial for the development of personalized treatment in diabetes care. The potential benefits include decreasing the risk of complications and reducing healthcare costs by allocating resources to patients who would benefit most.

The availability of large datasets and machine learning (ML)-based models trained on large health datasets has enabled the development of various prognostic tools for disease onset and diabetes-related adverse outcomes at the population level [7–11]. Clustering is one of the most useful methods for patient segmentation and analyzing patient similarities for precision medicine, thus various clustering methods have been utilized for the prediction of T2D risk or undiagnosed disease [8, 9, 12–14]. However, studies that aim to predict disease progression are still scarce. Ahlqvist’s classification of diabetes into five classes with different risks of diabetic complications is the most validated disease progression model to date [15–18]. However, not all of the six variables used for clustering, such as HOMA-2 indices, are routinely available in clinical practice, limiting their usability [7, 18].

The public healthcare system and comprehensive national health and social care registers in Finland offer excellent tools to study diabetes in a comprehensive, population-based patient population. The aim of this study was to characterize Finnish patients with T2D, their treatment patterns and disease progression utilizing a nationwide cohort with comprehensive data collection from over 300,000 Finnish patients. In addition, a data-driven clustering method was utilized to identify and characterize clinically meaningful patient clusters with different disease progression based on the occurrence of T2D-related complications.

Methods

Study Design

This was a nationwide, non-interventional, retrospective study utilizing data from the Finnish health and social care registers. The inclusion period was between January 1, 2010, and December 31, 2019. The observation period, including data collection from the registries, spanned from January 1, 1996 (or later, depending on the data availability of each register), to December 31, 2021, allowing a minimum of 2 years of potential follow-up for each patient (Fig. 1).

Fig. 1.

Overview of the study design and formation of the total type 2 diabetes (T2D) patient population. ATC Anatomical Therapeutic Chemical classification, ICD-10 International Classification of Diseases, Tenth Revision, ICPC-2 International Classification of Primary Care, 2nd edition

Study Population and Follow-Up

The study population consisted of adult patients (≥ 18 years of age) with incident T2D who fulfilled ≥ 1 of the inclusion criteria presented in Fig. 1 during 2010–2019, and none during the baseline period of 1996–2009.

The follow-up started at the first occurrence (index date) of any inclusion criterion and finished at the end of the observation period, or death, whichever came first. In annual analyses, the year before the index date and the subsequent follow-up years were treated separately. Patients contributed to the annual analyses if they were alive in the beginning of the year analyzed.

As common practice in machine learning, the study population was divided into 80% training and 20% test sets for the later stage of predictive modeling. In this study, baseline and follow-up characteristics and medication use are reported for the total population, and all further results for the training set. The current work serves as the basis for future modeling.

Data Sources and Variables

Detailed descriptions of the used data sources and collected variables are given in Table S1 and Table S2, respectively. Data permits for the study were granted by the Health and Social Data Permit Authority Findata and Statistics Finland (socioeconomic variables).

Characterization of the Study Population

For the total population, the demographics, socioeconomic status, clinical characteristics and medication use (other than diabetes medications) were described at baseline and during follow-up. Description of comorbidities utilized the entire baseline period, i.e., all data prior to the index date, and the laboratory values used were those from the last measurement prior to the index date. Medication use was reported from a 6-month period prior to the index. More detailed description of laboratory tests, diagnostic (the International Classification of Diseases, 10th revision [ICD-10], International Classification of Primary Care, 2nd edition [ICPC-2]), procedure (Nomesco Classification of Surgical Procedures [NCSP]) and Anatomical Therapeutic Chemical (ATC) codes collected during the study is given in Table S3. Reference values for the collected laboratory values are listed in Table S4. For the total population and the training set, the use of antidiabetic medications was reported for the entire follow-up, both as individual medications (ATC codes listed in Table S3) and further divided into the following four categories: 1) glucose-lowering treatment (including biguanides, dipeptidyl peptidase 4 [DPP-4] inhibitors, insulins [long-lasting, intermediate-acting, fast-acting, and for inhalation], sulfonylureas, thiazolidinediones, alpha-glucosidase inhibitors, and sulfonamides); 2) outcome-improving treatment (including glucagon-like peptide-1 [GLP-1] agonists and sodium/glucose cotransporter 2 [SGLT2] inhibitors); 3) combination of both glucose-lowering and outcome-improving treatment; and 4) no treatment. In addition, medication use for the four aforementioned categories was reported for 1-year time windows after the index date.

Disease Progression

For the training set, disease progression during the follow-up was defined by the presence of 19 clinical end-points, along with three additional healthcare resource utilization (HCRU), benefit and sick leave cost end-points. The 19 clinical end-points, which were primarily based on complications described in Diabetes in Finland (FinDM) project [19], are defined in detail in Table 1. HCRU cost was defined as belonging to the highest cost decile per time window, where the decile was determined across all the time windows. The social benefit end-point was defined as receiving ≥ 1000 euros in allowances from sick leave, rehabilitation benefits, and disability pensions during the annual time windows. The sick leave end-point was based on the total number of short sick leaves (each lasting < 10 days individually), with a threshold of 5 weekdays per annual time window. The presence of each of the predefined 22 end-points was assessed within the annual windows, and disease progression was described as a sum of these individual components.

Table 1.

Definitions for the 19 selected clinical end-points associated with type 2 diabetes progression

Cardiovascular complications
Acute myocardial infarction	≥ 1 of the following diagnoses ICD-10: I21, I46, R96, R98
Atrial fibrillation	ICD-10: I48
Cerebrovascular complications (ischemic stroke)	≥ 1 of the following diagnoses or procedure codes ICD-10: I63.0–I63.5, I63.7–I66.9, I67.2, I69.3, G45; NCSP: PA, PB
Coronary artery disease	≥ 1 of the following diagnoses or procedure codes ICD-10: I20-25; NCSP: FNA-FNE, FN1AT, FN1BT, FN1ST, FN1YT, FN2, FN3CT, TFN40, TFN50
Heart failure	ICD-10: I50, ICPC-2: K77
Hypertension	≥ 1 of the following diagnoses ICD-10: I10–I15
Peripheral artery disease	> 1 of the following diagnoses or procedure codes ICD-10: I70.2, I79.2, E11.5, I79.2 + E11.5; NCSP: PD–PG
Valve disorders	≥ 1 of the following diagnoses ICD-10: I34–I37, I39.0–I39.4
Kidney complications
Dialysis	≥ 1 of the following diagnosis or procedure codes ICD-10 Z49, NCPC: TK800, TK820
Diabetic kidney disorder	Laboratory value nUAlbKrea or U-AlbKrea > 3 mg/mmol OR eGFR < 60 ml/min/1.73 m2 with at least two positive findings separated by no more than 6 months OR ICD-10: E11.2 (on its own) OR ICD-10: E11.2, N08.03, Z49, Z94
End-stage renal disease	Laboratory value eGFR < 15 ml/min/1.73 m2 with at least two measurements separated by at least 30 days but not more than 12 months OR ICD-10: N18/Z99.2 OR a kidney transplant defined as ≥ 1 of the following diagnosis or procedure codes (inpatient or outpatient): ICD-10: Z94.0, T86.1; NCSP: KAS00, KAS10, KAS40 OR ≥ 1 dispensation with the following ATC codes: B03XA01, V03AE07, V03AE04, V03AE03, V03AE02, V03AE05, V03AE08
Other complications
Acute complications	Complications requiring hospitalizations: ketoacidosis and hyperglycemias, hypoglycemias, hyperosmolar hyperglycemic syndrome, ICD-10: E11.00, E11.01, E11.09, E11.1
Diabetic eye complications	≥ 1 of the following diagnoses or procedure codes ICD-10: E11.3, H28.0, H34, H36.0, H40.5, H42.0, H43.1, H45.0, H54; NCSP: CKC12, CKD05, CKD60, CKD65, or CKC*, CKD*
Diabetic circulatory complications	ICD-10: E11.5 (on its own)
Diseases of the limbs	≥ 1 of the following diagnoses or procedure codes ICD-10 (foot ulcer, infections, orthopedic problems caused by malposition of the foot and neuropathy, amputations): L97, L03.1, M86.1–M86.4, M86.6–M86.9, S91, T93; B35.1, B07, B35.3, L28, L85.3, L84, L03.0; M72.2, L60.0, M21.5, M21.4, M20.4, M20.2, M20.1; M14.6 + A52.0; Z89.7, Z89.6, Z89.5, Z89.4; NCSP: NFQ10, NFQ20, NGQ10, NGQ20, NHQ10, NHQ20, NHQ30, NHQ40, NHK99, NHS20, NHS99, NHW, QDB05, QDG20, QDG30, QDG99 Skin ulcerations in primary care ICPC-2: S97
Neuropathy	Diabetic neuropathy: ≥ 1 of the following diagnoses ICD-10: M14.2, M14.6, G57.3, G57.5, G59.0, G63.2, G73.0, G99.0, H49, I95.1, N48.4, E11.6, E11.5, E11.4
Dyslipidemias	ICD-10: I70
Other types of eye complications	Other types of eye complications (including proliferative retinopathy and macular swelling), blindness: ICD-10: H35*, H54
Mortality
All-cause mortality	Death was defined as death by any cause, and as yes/no per time window

eGFR estimated glomerular filtration rate, ICD-10 the International Classification of Diseases, 10th revision, ICPC-2 International Classification of Primary Care, 2nd revision, NCSP NOMESCO Classification of Surgical procedures

Identification of Patient Clusters

Patient clusters were identified in the training set based on the longitudinal disease progression patterns including the 22 end-points described above. The clustering used likelihood-based growth mixture modeling (GMM), which fits a specific number of latent growth trajectories to the data and then evaluates the posterior probabilities of each individual belonging to a specific trajectory group [20]. The individuals are assigned to the latent group in which their posterior probability is maximized, and these form the patient clusters. The clustering was performed using a pre-specified number of latent classes (3–5) and linear trajectories with random intercept and random slope. Models with 1 and 2 latent classes were performed for comparison. Model performance was evaluated using the Akaike information criterion (AIC) and Bayesian information criterion (BIC), average posterior probabilities, entropy, and a clinical assessment of the relevance of the disease progression patterns identified within each cluster.

Characterization of the Patient Clusters

The demographics, socioeconomic variables and clinical characteristics of the identified clusters were described at baseline and during follow-up, as described previously for the total population. Disease progression was shown both as the mean number of all end-points and the rate of individual end-points in the annual time windows. The characterization serves as pre-screening of independent variables that differentiate across the clusters and may be used to predict in which cluster patients belong to.

General Statistical Considerations

In descriptive analyses the data for continuous variables were summarized by presenting the mean with standard deviation. For categorical variables, the counts and proportions of each value were reported. Presented sample sizes are the number (n) of patients with non-missing values. P-values reported for the descriptive cluster comparisons were calculated with ANOVA for continuous variables and with the chi-square test for categorical variables. The analyses were conducted using R, version 4.2.1, and GMM was applied with the lcmm package.

Ethical Considerations

This study was approved by the Finnish data Permit Authority, Findata. No ethics approval or informed consent from the patients included in the study cohort was required by the Finnish legislation (Act on the Secondary Use of Health and Social Data (552/2019) by the Ministry of Social Affairs and Health), as the persons were not contacted, and the study did not affect the treatment of patients. The study was conducted in accordance with the Helsinki Declaration of 1975, Good Pharmacoepidemiology Practices, and Data Protection Directive.

Results

Baseline Characteristics of the Total Study Population

Description of the baseline sociodemographic characteristics, selected comorbidities, and selected laboratory values of the total population (N = 302,987) are presented in Table 2. Mean age of the population at diagnosis was 62.0 (SD 14.0) years, 48.3% of the study population were women and 55.9% pensioners. The most common baseline comorbidities were hypertension (38.7%), coronary artery disease (14.6%), and atrial fibrillation (10.6%). Depression was present in 12.5% of the patients. Clinical characteristics of the total population during the entire follow-up are presented in Table S6.

Table 2.

Sociodemographic and clinical characteristics of the total study population and the patient clusters (training set, 80% of the total study population; n = 242,390) at baseline (i.e., recording at or closest to index date)

	Total population (N = 302,987)	Cluster 1 (n = 133,951)	Cluster 2 (n = 52,819)	Cluster 3 (n = 43,488)	Cluster 4 (n = 10,159)	Cluster 5 (n = 1973)
Sociodemographic characteristics
Age (years), mean (SD)	62.0 (14.0)	58.3 (13.4)	62.9 (13.0)	68.3 (13.2)	74.9 (11.4)	78.6 (11.2)
Sex (women), N (%)	146,405 (48.3%)	65,893 (49.2%)	25,477 (48.2%)	20,102 (46.2%)	4589 (45.2%)	933 (47.3%)
Socioeconomic status, N (%)
Employed	102,109 (33.7%)	58,549 (43.7%)	14,849 (28.1%)	7548 (17.4%)	751 (7.4%)	100 (5.1%)
Long-term unemployed	21,212 (7.0%)	11,427 (8.5%)	3350 (6.3%)	1875 (4.3%)	224 (2.2%)	41 (2.1%)
Pensioner	168,385 (55.9%)	57,884 (43.2%)	32,959 (62.4%)	33,004 (75.9%)	9045 (89.0%)	1810 (91.7%)
Other	9315 (3.1%)	5047 (3.8%)	1400 (2.7%)	858 (2.0%)	102 (1.0%)	NA
Missing socioeconomic status	1966 (0.6%)	1044 (0.8%)	261 (0.5%)	203 (0.5%)	37 (0.4%)	NA
End-point comorbidities
Acute myocardial infarction, N (%)	15,768 (5.2%)	2816 (2.1%)	2546 (4.8%)	4885 (11.2%)	1988 (19.6%)	340 (17.2%)
Atrial fibrillation, N (%)	32,122 (10.6%)	4229 (3.2%)	4716 (8.9%)	11,029 (25.4%)	4823 (47.5%)	823 (41.7%)
Ischemic stroke, N (%)	27,594 (9.1%)	6281 (4.7%)	4593 (8.7%)	7675 (17.6%)	3064 (30.2%)	468 (23.7%)
Coronary artery disease, N (%)	44,144 (14.6%)	8813 (6.6%)	7706 (14.6%)	12,935 (29.7%)	4925 (48.5%)	852 (43.2%)
Heart failure, N (%)	15,881 (5.2%)	1534 (1.1%)	1579 (3.0%)	5351 (12.3%)	3501 (34.5%)	697 (35.3%)
Hypertension, N (%)	117,229 (38.7%)	36,413 (27.2%)	22,853 (43.3%)	26,088 (60.0%)	7135 (70.2%)	1172 (59.4%)
Peripheral artery disease, N (%)	6664 (2.2%)	730 (0.5%)	691 (1.3%)	2047 (4.7%)	1619 (15.9%)	278 (14.1%)
Valve disorders, N (%)	9195 (3.0%)	1138 (0.8%)	1205 (2.3%)	3014 (6.9%)	1697 (16.7%)	291 (14.7%)
End-stage renal disease, N (%)	2940 (1.0%)	194 (0.1%)	209 (0.4%)	832 (1.9%)	938 (9.2%)	145 (7.3%)
Chronic kidney disease, N (%)	2788 (0.9%)	189 (0.1%)	201 (0.4%)	775 (1.8%)	890 (8.8%)	138 (7.0%)
Dyslipidemias, N (%)	10,537 (3.5%)	2138 (1.6%)	1351 (2.6%)	2889 (6.6%)	1798 (17.7%)	297 (15.1%)
Other comorbidities
Depression, N (%)	37,915 (12.5%)	15,092 (11.3%)	7674 (14.5%)	6302 (14.5%)	1067 (10.5%)	137 (6.9%)
Anxiety, N (%)	20,453 (6.8%)	8517 (6.4%)	4005 (7.6%)	3272 (7.5%)	481 (4.7%)	64 (3.2%)
Acute stress reactions, N (%)	17,072 (5.6%)	7734 (5.8%)	3140 (5.9%)	2406 (5.5%)	316 (3.1%)	40 (2.0%)
Schizophrenia, N (%)	11,897 (3.9%)	4042 (3.0%)	2973 (5.6%)	2203 (5.1%)	251 (2.5%)	48 (2.4%)
Dementia, N (%)	6896 (2.3%)	1706 (1.3%)	963 (1.8%)	1963 (4.5%)	707 (7.0%)	196 (9.9%)
COPD, N (%)	11,356 (3.7%)	2839 (2.1%)	1900 (3.6%)	2951 (6.8%)	1185 (11.7%)	234 (11.9%)
Asthma, N (%)	29,882 (9.9%)	10,601 (7.9%)	5398 (10.2%)	5811 (13.4%)	1730 (17.0%)	267 (13.5%)
Laboratory measurementsa
B-HbA1c (mmol/mol), mean (SD)	40.8 (16.5)	42.4 (17.3)	39.1 (16.0)	39.5 (15.2)	40.3 (15.6)	40.8 (16.2)
B-HbA1c (%)b	5.9	6.0	5.7	5.8	5.8	5.9
Normal (< 42 mmol/l), N (%)	31,574 (10.4%)	11,663 (8.7%)	6537 (12.4%)	5621 (12.9%)	1188 (11.7%)	141 (7.1%)
Missing, N (%)	243,746 (80.4%)	110,407 (82.4%)	41,435 (78.4%)	33,482 (77.0%)	7944 (78.2%)	1697 (86.0%)
Fasting glucose (mmol/l), mean (SD)	7.7 (2.4)	7.8 (2.5)	7.7 (2.4)	7.6 (2.2)	7.5 (2.7)	7.7 (2.7)
Normal (≤ 6 mmol/l), N (%)	6354 (2.1%)	2295 (1.7%)	1096 (2.1%)	1233 (2.8%)	387 (3.8%)	59 (3.0%)
Missing, N (%)	238,563 (78.7%)	108,730 (81.2%)	40,155 (76.0%)	32,475 (74.7%)	7770 (76.5%)	1665 (84.4%)
Cholesterol (mmol/lc), mean (SD)	5.0 (1.2)	5.2 (1.1)	5.1 (1.2)	4.8 (1.2)	4.4 (1.2)	4.2 (1.2)
Normal (< 5 mmol/l), N (%)	23,618 (7.8%)	7399 (5.5%)	4683 (8.9%)	5145 (11.8%)	1438 (14.2%)	188 (9.5%)
Missing, N (%)	256,005 (84.5%)	116,710 (87.1%)	43,444 (82.3%)	34,822 (80.1%)	8110 (79.8%)	1735 (87.9%)
Triglycerides (mmol/lc), mean (SD)	1.9 (1.6)	1.9 (1.6)	1.9 (1.5)	1.8 (1.3)	1.752 (1.5)	1.7 (2.5)
Normal (< 1.7 mmol/l), N (%)	26,260 (8.7%)	9348 (7.0%)	5137 (9.7%)	5049 (11.6%)	1277 (12.6%)	173 (8.8%)
Missing, N (%)	257,648 (85.0%)	117,468 (87.7%)	43,776 (82.9%)	35,032 (80.6%)	8147 (80.2%)	1737 (88.0%)
LDL-cholesterol (mmol/lc), mean (SD)	3.1 (1.0)	3.3 (1.0)	3.1 (1.0)	2.9 (1.0)	2.6 (1.0)	2.8 (1.0)
Normal (< 3 mmol/l), N (%)	24,190 (8.0%)	7223 (5.4%)	4834 (9.2%)	5454 (12.5%)	1573 (15.5%)	195 (9.9%)
Missing, N (%)	251,408 (83.0%)	115,112 (85.9%)	42,528 (80.5%)	33,871 (77.9%)	7903 (77.8%)	1716 (87.0%)
HDL-cholesterol (mmol/lc), mean (SD)	1.3 (0.4)	1.4 (0.4)	1.4 (0.4)	1.3 (0.4)	1.3 (0.4)	1.3 (0.4)
Normal (> 1 (M); > 1.2 (F) mmol/l), N (%)	32,759 (10.8%)	12,210 (9.1%)	6619 (12.5%)	5867 (13.5%)	1343 (13.2%)	145 (7.3%)
Missing, N (%)	257,495 (85.0%)	117,357 (87.6%)	43,724 (82.8%)	35,063 (80.6%)	8162 (80.3%)	1742 (88.3%)
Creatinine (μmol/l), mean (SD)	75.9 (35.0)	70.2 (19.4)	73.3 (19.8)	78.7 (29.0)	106.9 (94.5)	112.8 (98.6)
Normal (50–95 mmol/l), N (%)	54,246 (17.9%)	20,032 (15.0%)	10,787 (20.4%)	10,339 (23.8%)	2008 (19.8%)	253 (12.8%)
Missing, N (%)	235,619 (77.8%)	110,682 (82.6%)	39,859 (75.5%)	29,810 (68.5%)	6679 (65.7%)	1462 (74.1%)
GFRed (ml/min/1.73 m2), mean (SD)	87.5 (29.2)	93.8 (29.0)	87.9 (25.9)	82.4 (28.5)	67.0 (28.6)	62.3 (26.4)
Normal (> 60 ml/min/1.73 m2), N (%)	61,393 (20.3%)	23,266 (17.4%)	12,040 (22.8%)	11,452 (26.3%)	2114 (20.8%)	270 (13.7%)
Missing, N (%)	233,621 (77.1%)	109,608 (81.8%)	39,616 (75.0%)	29,568 (68.0%)	6651 (65.5%)	1460 (74.0%)
Albumin (μg/min), mean (SD)	29.0 (165.9)	7.4 (23.0)	22.5 (87.4)	39.5 (141)	140.9 (448)	94.8 (206)
Normal (< 20 μg/min), N (%)	1527 (0.5%)	535 (0.4%)	395 (0.7%)	246 (0.6%)	37 (0.4%)	< 5
Missing, N (%)	301,209 (99.4%)	NA	52,359 (99.1%)	43,166 (99.3%)	10,103 (99.4%)	NA
Urine albumin-to-creatinine ratio (UACR) (mg/min), mean (SD)	7.3 (36.2)	2.1 (11.8)	5.1 (25.5)	12.7 (49.5)	35.3 (83.9)	68.2 (124.8)
Normal (< 3 mg/min), N (%)	3241 (1.1%)	1434 (1.1%)	554 (1.0%)	502 (1.2%)	88 (0.9%)	9 (0.5%)
Missing, N (%)	299,011 (98.7%)	132,382 (98.8%)	52,165 (98.8%)	42,726 (98.2%)	9971 (98.1%)	NA

B-HbA1c glycated hemoglobin, COPD chronic obstructive pulmonary disease, GFRe estimated glomerular filtration rate, HDL high-density lipoprotein, IFCC Internal Federation of Clinical Chemistry, LDL low-density lipoprotein, NGSP National Glycohemoglobin Standardization Programme

^aProportions of patients with normal / missing values denote the percentage in total population (N = 302,987)

^bConverted from the mean mmol/mol (IFCC) to % (NGSP) using a HbA1c calculator (https://ngsp.org/convert1.asp)

^cTo convert lipid levels from mmol/l to mg/dl, multiply by specific factors based on the lipid type: 38.67 for cholesterol (total, HDL, LDL) and 88.57 for triglycerides

^dPresented eGFR values are calculated using the CKD-EPI equation, based on patients’ plasma creatinine values, age, and sex

Medication Use During Follow-Up

During the entire follow-up, 81.2% (n = 245,909) of the total population used glucose-lowering drugs, 19.2% (n = 58,075) outcome-improving drugs (GLP-1 analogues and SGLT2 inhibitors), and 18.9% (n = 57,277) used both (Fig. 2A). The most used individual glucose-lowering drugs were biguanides (75.8%, n = 229,520), and outcome-improving treatments SGLT2 inhibitors (17.9%, n = 54,343) (Fig. 2B). Approximately one in five patients (18.6%, n = 56,280) had no pharmacy dispensations for T2D medication during the entire follow-up period.

Fig. 2.

Use of type 2 diabetes related medications in the total patient population (N = 302,987) during the entire follow-up. Medication use is presented both A grouped as glucose-lowering treatment, outcome-improving treatment, combination treatment, including both glucose-lowering and outcome-improving medication and no treatment and B as individual medications. DPP-4 dipeptidyl peptidase 4, GLP-1 glucagon-like peptide-1, SGLT2 sodium/glucose cotransporter 2

For indications other than diabetes, the most used medications were agents acting on the renin–angiotensin system (44.1% at baseline; 68.1% during follow-up), lipid modifying agents (32.3% at baseline; 65.8% during follow-up), analgesics (15.8% at baseline; 64.3% at follow-up) and beta blockers (36.9% at baseline; 53.1% during follow-up) (Figure S1).

Identification of the Patient Clusters

Five distinctive patient clusters were identified by applying the GMM on long-term disease progression data. The selection of five clusters over three or four was supported by both the AIC and BIC and ascertained by clinical evaluation (Table S5). Namely, with three or four clusters (Figure S2), progression patterns were not as clear as with five (Fig. 3).

Fig. 3.

Mean number of type 2 diabetes (T2D) related end-point complications in the five identified T2D patient clusters over the first 5 years of follow-up (clusters 1–5) in the training set (n = 242,390). The mean number of complications during the baseline year was 0.2 per patient in cluster 1, 0.5 in cluster 2, 1.2 in cluster 3, 2.3 in cluster 4 and 1.4 in cluster 5

Cluster 1 (C1: 44.2%, n = 133,951), cluster 2 (C2: 17.4%, n = 52,819) and cluster 3 (C3: 14.4%, n = 43,488) were the most common and included patients with either stable, mildly or modestly progressing disease (Fig. 3). Patients with rapid or extremely rapid long-term disease progression were allocated to cluster 4 (C4: 3.55%, n = 10,159) or cluster 5 (C5: 0.65%, n = 1973) that together comprised in total 4.2% of the study population. The number of patients in each cluster per follow-up year is shown in Figure S3.

During the baseline year, the mean number of complications increased gradually from C1 to C4; patients in C1 had 0.2 complications per patient, increasing to 0.5 in C2, 1.2 in C3 and 2.3 in C4. Patients in C5 had 1.4 complications per patients (Fig. 3). In C1–C3, the annual rate of complications was rather stable over the first five follow-up years, whereas in C5, the mean number of complications increased rapidly to 5.5 already during the first follow-up year, and further to 7.2 during the third follow-up year. A more modest increase was observed for C4 (Fig. 3).

Baseline Characteristics of the Five Patient Clusters

The baseline sociodemographic characteristics, selected comorbidities, and laboratory values of the five clusters are presented in Table 2. Age distribution of the patients moved towards higher age with severity of disease progression in the clusters, hereafter referred to as cluster severity (Figure S4). The proportion of patients with CVD was highest in clusters with more severe progression at baseline (Table 2): heart failure was present in only 1.1% and 3.0% of the patients in C1 and C2, respectively, in comparison to 34.5% and 35.4% of the patients in C4 and C5. Similarly, coronary artery disease was present in 6.6% in C1 compared to 48.5% in C4 and 43.2% in C5. On the contrary, depression (6.9%) and anxiety (3.2%) were less common in C5 than in other clusters (10.5–14.4% and 4.7–7.6%, respectively). Clinical characteristics of the clusters during the entire follow-up are presented in Table S6.

Medication Use in the Five Patient Clusters at Follow-Up

The percentage of patients using any medication for T2D during the entire follow-up decreased with cluster severity (Fig. 4A). In C1–C3, 76.5–83.3% of the patients were using glucose-lowering medication and 22.1–16.1% outcome-improving medication, compared to the respective proportions of only 68.1% and 8.5% in C4, and 51.5% and 1.4% in C5. Almost third of the patients in C4 (31.4%) and half of the patients in C5 (48.2%) did not use any T2D medication during the entire study follow-up (Fig. 4A).

Fig. 4.

Use of type 2 diabetes related medications in the training set (80% of the total study population; n = 242,390) during the entire follow-up A grouped as glucose-lowering treatment, outcome-improving treatment, combination treatment, including both glucose-lowering and outcome-improving medication and no treatment and B as individual medications. In cluster 5, < 5 patients using GLP-1 analogues and other glucose-lowering medications were recorded and thus not included to the graph. DPP-4 dipeptidyl peptidase 4, GLP-1 glucagon-like peptide-1, SGLT2 sodium/glucose cotransporter 2

The most used medication for T2D in the total population, biguanides, were used by 68.3–79.4% of patients in C1–C3, but only 51.3% in C4 and 35.3% in C5. Of the outcome-improving treatments, patients in C1 and C2 used the most SGLT2 inhibitors (19.0–20.6%) and GLP-1 analogues (4.6–6.1%). The use of other diabetes medications generally decreased with cluster severity, only the percentage of patients using long-lasting insulin increased from 6.4% in C1 to 10.5% in C4 and 8.1% in C5 (Fig. 4B).

The most frequently used medications for other indications than diabetes at follow-up were for CVD (agents acting on the renin–angiotensin system 51.0–79.3%, beta blockers 38.6–84% and lipid modifying agents 41.2–72.2%, depending on cluster) and analgesics (44.6%–76.0%; Figure S5). Interestingly, the percentages of patients using any of these medications were lower in C1 (best prognosis) and C5 (worst prognosis), than in C2–C4 (Figure S5).

Annual Rate of Clinical End-points in the Total Population and Patient Clusters

The annual rate of the complication end-points for both the total population and for the five clusters are shown in Fig. 5 for coronary artery disease, heart failure, cerebrovascular complications, acute complications, end-stage renal disease and mortality, and in Figure S6 for other end-points.

Fig. 5.

Annual rate of A coronary artery disease, B heart failure, C mortality, D end-stage renal disease, E ischemic stroke and F acute complications in the type 2 diabetes patient clusters (training set, 80% of the total study population; n = 242,390). Figures indicate the annual rate of each complication end-point during the respective follow-up year (year 0, year 1 etc.), i.e., they are not cumulative during the years and thus, do not represent the overall prevalence of complication end-points in the total patient population or in the identified patient clusters. Year 0 refers to the baseline year before the index date, year 1 to the first follow-up year, and so on

In the total population, the annual rate of all CV end-points increased until approximately the 10th year of follow-up. Similarly, the annual rate of all kidney complications increased steadily with the follow-up years. A visible peak in the annual rate of most clinical end-points was observed during the first follow-up year. The peak was the most prominent in acute complications (complications requiring hospitalizations including ketoacidosis and hyperglycemias, hypoglycemias, hyperosmolar hyperglycemic syndrome). Yearly mortality in the total population was highest during the first year after the diagnosis (2.8%), after which it varied between 2.0–2.6%.

Clear differences in the annual rate of complication end-points were observed between the five patient clusters. Coronary artery disease (Fig. 5A), heart failure (Fig. 5B) and cerebrovascular complications (Fig. 5E) increased rapidly already during the first 2 years in C5 and clearly during the first 4 years in C4. On the contrary, in clusters C1–C3 the overall increase in annual rate of the CV complications was modest. In C5, mortality was remarkably high: over 75% patients died already during the first follow-up year and all by the 5th follow-up year (Fig. 5C). ESRD increased in all clusters during the follow-up, and the most significant increase was observed in C4 and C5 (Fig. 5D). The annual rate of acute complications was highest during the first follow-up year and decreased thereafter in all clusters (Fig. 5F).

Screening of Explanatory Variables for Early Identification of Patient Clusters

Selected explanatory variables that could identify the cluster a patient belongs to, based on the observed differences across clusters at baseline and during the first follow-up year, are presented in Fig. 6. The annual rate of patients with certain infections and abnormal values of creatinine, hemoglobin, and erythrocytes increased with cluster severity. The use of antithrombotics, beta blockers, diuretics, heart medications, and renin-angiotensin agents was more common in severe clusters during the first year of follow-up. In addition, the proportion of patients without treatment for T2D was higher in severe clusters: 23.9% in C1 and 35.3% in C5.

Fig. 6.

Annual rate of A antithrombotics, B beta blockers, C diuretics, D heart medications, E agents acting on the renin-angiotensin system, F) certain infections and parasitic diseases (ICD-10: A*), G acute pyelonephritis, cystitis, other disorders of urinary system, H COPD, I glucose hemoglobin, J hemoglobin, K creatinine, L erythrocytes, M glucose-lowering treatment, N outcome-improving treatment, O both treatments, and P no treatment in the five type 2 diabetes patient clusters (training set, 80% of the total study population; n = 242,390). Figures indicate the annual rate of each complication end-point during the respective follow-up year (year 0, year 1 etc.), i.e., they are not cumulative during the years and thus, do not represent the overall prevalence of complication end-points in the total patient population or in the identified patient clusters. Year 0 refers to the baseline year before the index date, year 1 to the first follow-up year, and so on. ICD-10 International Classification of Diseases, Tenth Revision

Discussion

In this nationwide register study, we characterized treatment patterns and disease progression in a Finnish population including all incident patients with T2D between 2010–2021 and showed that patients can be allocated into five clinically meaningful clusters representing distinct disease progression profiles. At baseline, these five clusters can be considered to present mild (C1, C2) and moderately severe (C3, C4, C5) disease types. Moreover, their different longitudinal patterns during follow-up distinguished stable (C1), mild (C2), moderate (C3), rapid (C4), or extremely rapid (C5) progression patterns. Notably, the percentage of patients not using any T2D medication during follow-up increased with cluster severity from 16.4% in C1 to 48.2% in C5.

Our total study population consisted of over 300,000 incident patients with T2D, whose mean age at index date was 62.0 years, in line with a report from Sweden (62.9 years at diagnosis), but higher than what has been reported in the US (49.9 years at diagnosis) [21, 22]. In line with previous reports, cardiovascular comorbidities were common also in our cohort [22–24]. The prevalence of diagnosed CKD at baseline was very low (0.9%), however, when patients’ laboratory values were assessed to determine whether they met the criteria for CKD, the prevalence was found to be over ten times higher (10.5% of patients with eGFR available), indicating that the condition is underdiagnosed among patients with T2D, supporting previous findings [25–27]. Therefore, annual screening of eGFR and albuminuria, along with the implementation of evidence-based therapies such as GLP-1 analogues and SGLT-2 inhibitors, are essential to reduce the risk of disease progression and cardiovascular outcomes in these patients.

Artificial intelligence (AI) has been suggested to revolutionize personalized diabetes care by providing the means to both predict risk and tailor treatment plans based on the patient’s clinical presentation to improve outcomes [24, 28]. For prediction of the development of T2D, a recent study showed that biological factors, most notably HbA1c measured at baseline, surpassed the traditional risk factors, including diet, physical activity, and socioeconomical status for predicting disease risk within a 10-year time frame [14]. In addition, several AI/ML-based tools have already been developed for prediction of disease progression; however, they have so far most often been based on clinical and demographic characteristics of the patients [7, 29]. Here, we were able to identify five distinctive patient clusters, representing different long-term disease progression profiles from stable (C1) to extremely rapid (C5), based on the appearance of 22 selected end-points representing disease-related complications rather than biological variables. Interestingly, despite different long-term disease progression (presented in Fig. 3), clusters 3 and 5 had very similar number of end-point complications at baseline (1.2 and 1.4 end-point complications during the baseline year, respectively), whereas cluster 4 had the highest number of end-point complications (2.3 end-points). As previously suggested by Ahlqvist and colleagues, it would be reasonable to target the patients with severe disease progression with intensified treatment at diagnosis, to prevent future T2D-related complications [7].

A consensus report by the American Diabetes Association (ADA) and the European Association for the study of Diabetes (EASD) highlights the importance of early glucose-lowering treatment to attain recommended glycemic targets to reduce the onset and progression of microvascular complications [30]. The most used individual medications for T2D in our Finnish population were biguanides, including metformin. Biguanides were less used in C4 and C5 than in C1–C3, this could be explained by higher number of contraindications, such as renal insufficiency, in the severe clusters. Metformin use decreased during follow-up in line with a previous Finnish report [31], likely reflecting the need for treatment intensification as the disease progresses. The highest annual rate of acute complications (e.g., ketoacidosis and hypoglycemia) during the first follow-up year across all clusters could reflect a wrong diagnosis or failure of initial insulin therapy. The use of outcome-improving treatments was negligible during the first follow-up years, as these medications became available later during the inclusion period and increased only during the late follow-up years. To assess the association between the use of outcome-improving treatments and long-term disease progression, the follow-up should be restricted to the period when treatments were available.

Surprisingly, in our analysis based on medication dispensations of the total population, approximately 20% of the patients did not use any medication for T2D during the entire follow-up. Moreover, the number increased to over 30% of the patients in C4 and almost half of the patients in C5 (48.2%). Since our study relies of dispensations of reimbursed medications, it could be possible that C5 represents a patient population particularly non-adherent to a described treatment. Poor treatment coverage within diagnosed cases could be due to apprehension, poor financial and/or physical access to medicines, or lack of information and support to adhere to prescribed medications [32]. Moreover, it could be possible that these elderly patients have less strict glucose targets due to frailty, poor life control or other diseases. Higher age might also influence T2D medication in C4 and C5; clinical practice guidelines focus on avoiding hypoglycemia and other adverse effects and tend to have more relaxed HbA1c targets for elderly, frail or medically complex patients [33]. The patients could also have more inpatient periods, and they received medications via the hospital, and thus is not shown in our data based on dispensations or are not treated anymore due to poor prognosis. In summary, C4 and C5 represent elderly populations with high CVD burden and likely also frailty people at their end-of-life stage. Nevertheless, this finding underlines the need for improved methods for screening and identification of patients with poor prognosis to improve their access to early treatment and care.

In addition to identification of patient groups with different risks of poor prognosis, it would be highly beneficial to identify factors that could be used to allocate individual patients to their respective groups at an early disease state. Based on our findings, important explanatory variables to indicate poor prognosis could be cardiovascular comorbidities at or soon after T2D diagnosis. As expected based on the reported high mortality due to CVD among patients with T2D [24], individuals with diagnosed hypertension, coronary artery disease, or atrial fibrillation, and individuals using antithrombotics, beta blockers or renin-angiotensin agents were allocated to clusters with higher risk of poor prognosis. Even though HbA1c has been shown to be crucial for diabetes prediction models [14], it was not among the most important variables to indicate poor progression based on the selected complication end-points. Our results support the recommendations of the European Society of Cardiology and EASD 2019 guidelines that evaluation of patients’ long-term disease progression risk should not be solely based on markers related to glucose metabolism. Rather, a more generalized approach that would include attaining the recommended glycemic targets, evaluating individual CVD risks, and assuring optimal treatments of present CVD would be warranted [34].

The key question for any clustering analysis is the clinical utility of the findings. In clinical practice, an automated system to capture patients with the highest complication risk as early as possible using comprehensive data from the patient’s medical history would be valuable. The presented findings will be an initial step in our future work that aims to develop more refined AI prediction algorithms that could be used for early stratification of patients with T2D into distinct clusters based on their disease progression risk. This would allow more personalized treatment paths from an early disease stage, improving outcomes and allocating resources more efficiently.

The strength of this study is the large and comprehensive data set that represents all the more than 300,000 patients with incident diagnosis between 2010–2019 in Finland. However, this study is not without limitations. First, this was a retrospective study based on national register data, and missing variables cannot be ruled out. Data from private sector healthcare visits was obtained from only one health care provider (Terveystalo), thus the capture rate of private sector visits is below 100%. Second, medication use was based on dispensations, meaning that both patients without prescribed medication and patients with prescribed medication that was not acquired from the pharmacy would be classified to the no treatment group. In addition, the coverage of laboratory data includes only public healthcare from the Wellbeing Services County of Southwest Finland (Varha), Wellbeing Services County of Northern Savonia (PSHVA), HUS Group, the joint authority for Helsinki and Uusimaa, and private healthcare from Terveystalo.

Regarding the performance of the GMM, although each of the clusters is clinically distinguishable at the population level, C3 had the lowest mean posterior probability (67%). This indicates overlap in posterior membership across clusters, and individual patient cluster assignments may be uncertain in C3. Furthermore, disease progression and mortality were jointly modeled in the outcome, with varying follow-up lengths across participants. This may be reflected in the results of C5 in which the follow-up time was at most 5 years, mainly due to death. Consequently, C5 may present frail people at their end-of-life stage, potentially suffering from several other diseases as well. However, C5 is clinically interesting as their baseline status is seemingly good, while their progression towards death is rapid. This may represent individuals who have limited or delayed engagement with healthcare services and therefore only met care at a late stage.

Conclusions

In this study, we characterized a comprehensive Finnish cohort of over 300,000 patients with T2D and further identified five distinct and clinically meaningful patient clusters representing different disease progression profiles based on complication end-points. According to our results, diagnosis of several other new conditions, particularly CVD comorbidities, at or soon after incident T2D diagnosis indicates a poor prognosis. These findings support a comprehensive approach to assessing long-term diabetes risk, including CVD evaluation and treatment alongside glycemic control, and serve as an initial step to build AI models as tools for healthcare professionals for early identification of patients at risk for future complications, which would allow early treatment optimization.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

Aino Vesikansa, Mirkka Koivumaki, and Juha Mehtälä collected and analyzed the data. All authors developed the study concept and design. Aino Vesikansa and Juha Mehtälä drafted the manuscript. Kim Nygård, Riitta Ryhänen, Aaron Kortteenniemi, Kajsa Järvinen, Aino Vesikansa, Juha Mehtälä, Mirkka Koivusano, Annakaisa Tirronen, Eralda Asllanaj, Jarmo Kaukua, Sami Pakarinen and Juha Tuominen took part in interpreting the results, reviewed the manuscript critically for important intellectual content, and gave final approval of the version to be published.

Funding

This study and the Rapid Service Fee were funded by Boehringer Ingelheim Ky, Helsinki, Finland.

Data Availability

The datasets generated and analyzed during the current study are not publicly available, since according to Finnish legislation, access to individual-level data is restricted only to individuals named in the study permit. The study protocol is available upon reasonable request from the corresponding author.

Declarations

Conflict of Interest

Annakaisa Tirronen, Eralda Asllana, and Jarmo Kaukua are employees of Boehringer Ingelheim Ky and B.V. Aino Vesikansa, Mirkka Koivumäki, and Juha Mehtälä are employees of MedEngine Oy. Kim Nygård, Riitta Ryhänen, Aaron Kortteenniemi, Kajsa Järvinen, Sami Pakarinen, and Juha Tuominen have nothing to disclose. The authors meet the criteria for authorship as recommended by the ICMJE. The authors did not receive payment related to the development of the manuscript. Anniina Snellman, PhD, and Harlan Barker, MSc, of MedEngine Oy provided writing assistance, which was contracted and funded by Boehringer Ingelheim.​ Boehringer Ingelheim was given the opportunity to review the manuscript for medical and scientific accuracy as well as intellectual property considerations. The study was supported and funded by Boehringer Ingelheim.

Ethical Approval

This study was approved by the Finnish data Permit Authority, Findata. No ethics approval or informed consent from the patients included in the study cohort was required by the Finnish legislation (Act on the Secondary Use of Health and Social Data (552/2019) by the Ministry of Social Affairs and Health), as the persons were not contacted, and the study did not affect the treatment of patients. The study was conducted in accordance with the Helsinki Declaration of 1975, Good Pharmacoepidemiology Practices, and Data Protection Directive.