Forecasting Cancer Incidence Trends in Indian States and Union Territories (2019–2033): A Data-Driven Approach

Cancer burden in India

India’s NCRP predicts a rise in cancer cases, with 26.7 million DALYs caused in 2021 and 29.8 million by 2025, with northern and northeastern states being particularly affected, with men disproportionately burdened.^[6]

A study predicts a 12% increase in India’s cancer incidence over the next 5 years, with six regions showing significant differences. The most affected states are Madhya Pradesh, Tamil Nadu, Bihar, Maharashtra, Assam and Uttar Pradesh.^[7]

The study analysed cervical cancer cases from 2012 to 2015 using PBCRs. Ahmedabad Urban had the highest survival rate, but regional disparities were noted. The study emphasises awareness, early detection and healthcare access.^[7]

The study analysed cervical cancer cases from 2012 to 2015, revealing high survival rates in Ahmedabad Urban, but regional disparities, emphasising awareness, early detection and healthcare access.

Global prevalence of cancer

Study design

The trends in cancer incidence in Indian states and union territories (UT) from 2019 to 2023 were examined in this study using secondary data and a quantitative, retrospective design. The study employed both descriptive and inferential statistical methods to assess regional growth rates and forecast future incidence patterns. We developed forecasting algorithms to identify emerging cancer hotspots and inform policy suggestions.

Data source

Table 1 presents the past time series data of cancer growth rate. The information was sourced from credible public health databases and official government records, specifically: The ‘State/UT-wise Number of Estimated Incidence of Cancer Cases in India’ dataset (Unstarred Question No. 921, Session 265, 30 July 2024) was made available on data.gov. in and sent to the Rajya Sabha. The data are available on the website https://www.data.gov.in/resource/stateut-wise-number-estimated-incidence-cancer-cases-india-different-both-sexes-all-sites.

Data input

The input data have been obtained from the ‘State/UT-wise Number of Estimated Incidence of Cancer Cases in India’.

Trend analysis

On the basis of Table 1 and Figure 1, we have categorised the cancer growth rates in the states in the following categories:

• (i) States/UT with the relatively constant growth rate: Uttar Pradesh, Bihar, Odisha, Jharkhand, Madhya Pradesh, Gujarat, Karnataka, Tamil Nadu, West Bengal and Assam are showing a relatively constant growth rate in cancer.

• (ii) States/UT with increasing growth rate: In this category, we have observed Manipur, Mizoram, Pondicherry, Lakshadweep, Andaman and Nicobar Islands and Daman.

• (iii) States/UT with the declining pattern: In this study, we have observed some states/UT showing the declining pattern in the cancer growth, such as Andhra Pradesh, Telangana, Kerala, Goa, Ladakh, Chandigarh and Himachal Pradesh.

Close

Figure 1: Growth pattern of cancer incidence in India.

Export to PPT

India’s cancer incidence is predicted to rise by 12% over the next 5 years, with 29.8 million DALYs attributable by 2025, up from 26.7 million in 2021. The burden is disproportionately borne by northern and northeastern states, particularly for men.^[9]

West Bengal and Bihar show moderate growth rates in cancer cases, possibly due to improvements in detection, but also face challenges such as regional healthcare disparities and public health awareness. Smaller states such as Sikkim, Tripura and Meghalaya show erratic growth patterns, indicating potential challenges in maintaining stable data collection and cancer care reporting systems. Kerala, Tamil Nadu and Karnataka show relatively stable growth rates, with Kerala’s growth rate consistently low at 1.7–1.8% over 4 years, indicating a well-established healthcare infrastructure with effective cancer prevention and control programs.

Tools and software

Software used: R (version 4.3.1), which is an open-source statistical computing software that comes with enriched statistical libraries. We have used libraries ‘deplyr’, ‘forecast’ to forecast the cancer trends.

Microsoft Excel: A widely used software, Microsoft Excel is used to filter.

Identification of hotspots

The hot spot value is calculated by

The average growth rate over the 2019–2023 period was computed for each region to identify hotspots and trends.

The average growth rate is presented in Table 2.

To identify cancer hotspots, regions were rated based on their average growth rate over the 4-year period. According to Table 2 and Figure 2, the regions with the quickest growth rates were Daman and Dadra and Nagar Haveli. Sikkim, Manipur, Puducherry, Mizoram, Lakshadweep, Delhi and Chandigarh also had high development, albeit sporadically.

Close

Figure 2: Hot spot Regions for the cancer incidence

Export to PPT

The top five high-growth regions – Daman, Dadra and Nagar Haveli, Sikkim, Manipur and Mizoram – were compared to all other regions. As seen in Table 3, the average growth rate in these hotspots was significantly greater (5.65%) than the mean of 3.41% in other regions. The Top 5 group’s maximum increase was 8.10%, while the comparison group’s growth peaked at 3.60%.

Figure 2 depicts the average growth rates. The red bars show the states/UTs that have registered with the Daman and Dadra Nagar Haveli, which have reported the largest growth. Sikkim, Manipur, Puducherry, Mizoram, Lakshadweep, Delhi and Chandigarh have registered the higher growth rates.

The growth rate of cancer cases in India is significantly different between the five ‘hot-spot’ territories and the few ‘other’ regions. This result is presented in the Table 3. The top 5 states and UTs, led by Daman and Dadra and Nagar Haveli, are increasing their annual cases by approximately 5.7%, two-thirds faster than the comparison group’s 3.4% growth rate. The top -5 regions growth rate and the other regions growth rate are presented in the Table 3. The extremes of the disparity are even more evident, with Daman expanding at 8.1% and Mizoram surpassing all other regions in the ‘Others’ bucket. The lower median of 4.36% suggests that two high-growth western UTs are pushing the group’s mean upward, raising concerns about internal variability. The ‘Others’ are closely clustered between 3.05 and 3.60%, indicating lower pressure but a more uniform distribution. The top 5 states/ UTs with the higher growth rate alone are responsible for 62.36% of cases in India, making them priority targets for epidemiological research and resource distribution. Figure 3 represents the comparison of the Top -5 regions with high prevalence rate of cancer with the other regions.

Close

Figure 3: Comparison of the Cancer growth in the Top 5 states with the other.

Export to PPT

Forecast model

The study analysed the year-over-year percentage growth rate of cancer incidence in each state/UT from 2019 to 2023 to identify temporal changes, categorise states into low, medium and high-growth groups and identify accelerating or decelerating trends.

log(Y_t) = α + βt (1)

Where Y_t is the estimated number of new cancers α is the model intercept

β = growth rate coefficient

t = time.

To predict the incidence of cancer from 2024 to 2028, two forecasting models were used:

The lognormal distribution has been used to forecast the cancer incidences in the top States in India.

log(Y_t) = β₀ + β₁ * t (2)

Y_t = Predicted cancer count at year t

t = year of the incidence

β₀ = intercept

β₁ = slope

After fitting the model, the logY_t was transformed to the original Y_t

Y_t=e^(β0+β1*t) (3)

Selected states and UTs were subjected to these models:

• High-burden areas include Maharashtra and Uttar Pradesh.

• Regions with volatile growth: Daman and Sikkim

• Region with a low burden: Lakshadweep.

Model evaluation

Model performance was assessed using the following statistical metrics:

• R² (coefficient of determination): Indicates the proportion of variance explained by the model

• F-statistic and associated p-value: Test the overall significance of the regression

• Residual standard error (RSE): Measures the typical deviation of observed values from the fitted line

• 95% confidence intervals: Provided for all forecasts to account for uncertainty in estimates.

The final model demonstrated an excellent fit, with an R² value of 0.9998 and a statistically significant F-statistic, indicating high predictive reliability. Log-linear model R square and p value<0.05. The result s fitting is presented in the Table 6.

Limitations

• The model performs well in short-term forecasts, but its long-term accuracy is limited in a few ways:

• The model could only be trained on historical data spanning 5 years.

• Unaccounted confounders (e.g., migration, policy changes and health system improvements) may influence future incidence.

• The assumption of exponential growth should be reexamined when new data becomes available, as it may not always be accurate.

Manipur (2024–2033), comparing the four models:

The analysis of cancer incidence in five regions from 2024 to 2033 using log-linear regression (LOG), ARIMA and ETS models reveals a steady increase in cases. Daman’s incidence increased from 193.1 in 2024 to 459.0 by 2033, while Dadra’s increased from 146.7 to 438.4. The log-linear model predicts a dramatic rise in cancer cases, possibly due to better detection, increased risk exposure or population growth. ARIMA and ETS models provide more cautious predictions, with Daman and Dadra ending at 312.5 and 258.6, respectively. Manipur and Sikkim also show significant growth, with LOG models predicting they will double by 2033. ETS predictions are more conservative, with Mizoram having the lowest overall predicted incidence. The random forest model’s constant values and ARIMA’s flat projections suggest potential limitations in the models.

Four models were evaluated for forecasting accuracy: Log-linear, ARIMA, ETS and random forest. The log-linear model performed exceptionally well, with an R-squared of 0.9981 and a very low RMSE of 0.0076, indicating a nearly perfect fit with little error. Strong statistical reliability was confirmed by the model’s highly significant slope (β = 0.09622, p < 0.001) and intercept of −189.49.

With root mean square error (RMSE) = 1.9371, mean absolute error (MAE) = 1.5212 and a mean absolute percentage error (MAPE) of only 1.07%, the ARIMA (0, 1, 0) model also did well; its drift term (13.75 ± 1.08) and low autocorrelation (ACF1 = −0.103) point to a stable, objective model. With RMSE = 6.9903, MAE = 6.2497 and MAPE = 4.55%, the ETS (A, N and N) model demonstrated a moderate capacity for prediction.

These measures attest to the model’s strength and dependability in accounting for more than 99.8% of the variance in past cancer incidence data.

We have estimated the cancer growth rates for the states and UTs with the highest incidence trends using a log-normal distribution with a 95% confidence interval. Table 4 displays these projections for the years 2024–2033. It is crucial to remember that long-term projections should be interpreted cautiously due to the short duration of the historical data that are currently available. Figure 4 represents the growing trend of the cancer in the various regions by the selected models.

Close

Figure 4: Cancer incidence forecast for the period2024 -2033, by the models under study. ARIMA: Auto regressive integrated moving average, ETS: Exponential smoothing state space models, RF: Random forest.

Export to PPT

Model strength and fit

The study’s log-linear regression model fit the historical cancer incidence data exceptionally well, explaining more than 99.8% of the variance (R² = 0.9998). The model’s robustness in capturing growth patterns over time was demonstrated by its minimal RSE of 0.003955 and statistically significant F-statistic (45,060; p = 2.714e-16). The calculated coefficients, slope = 0.09244 and intercept = −180.9, show a steady increase in cancer incidence that is consistent with both national and international epidemiological trends.^[2,5]

The study^[9] highlights that short- to medium-term forecasts can be accurate and predictable, but models relying on small historical windows may experience decreased accuracy after 5 years due to unforeseen changes in the healthcare system, environment or socio-political landscape. Figure 5 shows the comparison of the fitted values vs the actual values. All the models are showing good fit graphically for the given period.

Close

Figure 5: Model forecast vs actual result. ARIMA: Auto regressive integrated moving average, ETS: Exponential smoothing state space models.

Export to PPT

Regional variation and forecasted growth

The study estimated the incidence of cancer in a number of high-growth states and UTs, including Sikkim, Manipur, Mizoram, Daman and Dadra and Nagar Haveli, using model forecasts and 95% confidence intervals. The incidence in Sikkim is predicted to increase by 49.6% over the next 10 years, from 541.95 in 2024 to 811.21 in 2033. The number of cases in Manipur, a region with volatile growth, is expected to increase from 2,274 in 2024 to 3,330.2 by 2033.

Despite having a small population, Dadra and Nagar Haveli have experienced rapid growth, rising from 274.35 in 2024 to 539.69 in 2033, indicating that it is a high-priority hotspot.

Daman also shows growth, with projected cases rising by more than 100% from 174.37 in 2024 to 362.06 in 2033.

These patterns are alarming, especially since these areas usually have inadequate infrastructure for tertiary cancer care. Healthcare systems in the northeast and small UTs are finding it difficult to handle the growing cancer burdens due to a shortage of oncology personnel, delayed diagnosis and a lack of preventive screening, according to Kulothungan et al. (2022) and Shafi et al. (2023).^[6,7]

Evaluation of models

Significant methodological insights were obtained from the comparative modelling utilising log-linear, ARIMA, ETS and random forest:

When it came to forecasting exponential growth, the log-linear model was the most aggressive. It is appropriate for short-term forecasts in patterns that increase steadily due to its high R² and low residual error. With an MAPE of 1.07% and an AIC of 21.53, the ARIMA (0, 1 and 0) model showed moderate sensitivity, indicating stable performance but limited ability to capture abrupt increases.

Forecasts from ETS (A, N and N) were smooth, but they were less accurate in volatile areas (MAPE = 4.55%). Possible under-fitting is indicated by its residual autocorrelation (ACF1 = 0.446).

In terms of trend detection, the random forest model performed poorly, generating forecasts that were flat across years. The RF model probably suffered in this case due to its low temporal granularity and lack of feature diversity, despite its effectiveness in high-dimensional non-linear datasets.

These findings corroborate the suggestion made by Desai et al. (2024)^[8] that, in epidemiological surveillance, simpler parametric models with solid theoretical foundations frequently perform better than black-box models when the volume of data is small or stable over time.

Implications for cancer control

The necessity of localised cancer control strategies is highlighted by the identification of the top five hotspots (Daman, Dadra and Nagar Haveli, Sikkim, Manipur and Mizoram), which have growth rates ranging from 3.7% to 8.1%, as opposed to 3.4% in other regions. Such disparities are associated with poor awareness, delayed health-seeking behaviour and unequal distribution of cancer care infrastructure, according to the NCRP (Indian Council of Medical Research and National Cancer registry Program[ICMR-NCRP]).

Moreover, the financial impact of cancer in high-growth areas poses another layer of concern. OOPE in states with inadequate public health funding frequently results in treatment discontinuation, which raises mortality and the long-term socioeconomic burden, as Prinja et al. (2023) pointed out.^[10]

Policy recommendations

Several suggestions can be made in light of these findings:

• To enhance data integrity and surveillance, fortify cancer registries in high-risk areas.

• Funding and infrastructure should be given priority to developing high-growth areas, particularly those with a shortage of tertiary care.

• Increase public health education and screening initiatives, particularly in the western and north eastern Utah regions, where delayed diagnosis is prevalent.

• As stressed by Schiliro and Firestein (2021),^[1] incorporate artificial intelligence-based early detection tools in high-incidence areas to support diagnostics.

Limitations

Despite the model’s excellent performance, some drawbacks must be noted:

• Five years of historical data were used for the analysis, which might not have adequately captured long-term cyclical patterns or changes brought about by policy.

• There was no modelling of confounding variables such as migration, occupational exposures, environmental pollution and modifications to screening procedures.

• The premise of exponential growth might not always be true, for example measures like tobacco control or HPV vaccination might flatten trends.