Source-linked AI summary
Deep learning via LSTM models for COVID-19 infection forecasting in India
Rohitash Chandra, Ayush Jain, Divyanshu Singh Chauhan
TL;DR
COVID-19 forecasting is hindered by unreliable models, incomplete data, and difficult-to-measure transmission factors. The paper applies LSTM-based models to Indian state data covering both major infection waves and forecasts infections across two months. It reports declining forecasts for October and November 2021, while emphasizing persistent uncertainty and modelling limitations.
Problem
COVID-19 forecasting remains difficult because existing models are unreliable and data are limited for representing transmission, reporting, population, travel, lockdown, and social factors.
Method
The paper evaluates univariate and multivariate LSTM-based models for Indian COVID-19 hotspot states using data covering the first and second infection waves.
Results
The univariate random-split ED-LSTM model provides the best test performance compared with the other evaluated models, while two-month forecasts show a steady decline in new cases.
Takeaways & Limitations
The framework supports short-term COVID-19 infection forecasting for India and can be applied to other countries and regions as more data become available.
Takeaways & Limitations
The framework cannot adequately capture social-cultural factors, population density, lockdown levels, and inter-state travel because of missing information and data.
Abstract
from arXiv · showhide
The COVID-19 pandemic continues to have major impact to health and medical infrastructure, economy, and agriculture. Prominent computational and mathematical models have been unreliable due to the complexity of the spread of infections. Moreover, lack of data collection and reporting makes modelling attempts difficult and unreliable. Hence, we need to re-look at the situation with reliable data sources and innovative forecasting models. Deep learning models such as recurrent neural networks are well suited for modelling spatiotemporal sequences. In this paper, we apply recurrent neural networks such as long short term memory (LSTM), bidirectional LSTM, and encoder-decoder LSTM models for multi-step (short-term) COVID-19 infection forecasting. We select Indian states with COVID-19 hotpots and capture the first (2020) and second (2021) wave of infections and provide two months ahead forecast. Our model predicts that the likelihood of another wave of infections in October and November 2021 is low; however, the authorities need to be vigilant given emerging variants of the virus. The accuracy of the predictions motivate the application of the method in other countries and regions. Nevertheless, the challenges in modelling remain due to the reliability of data and difficulties in capturing factors such as population density, logistics, and social aspects such as culture and lifestyle.
1 Introduction
COVID-19 created substantial health, economic, agricultural, and social disruption, while forecasting remained difficult because infection dynamics and reporting data are complex. The paper evaluates LSTM-based approaches for forecasting Indian COVID-19 infections across both major waves.
- 1 Introduction: The pandemic also caused widespread economic, agricultural, and social disruption through lockdowns, border closures, and effects on low-income communities and migrant workers.The reported global burden by October 2021 exceeded 238 million cases and 4.8 million deaths.
- 1 Introduction: India experienced two major infection waves, spanning April–October 2020 and February–June 2021, motivating forecasts that incorporate both periods.The first case in India was reported on 30 January 2020.
- 1 Introduction: Forecasting COVID-19 infections is difficult because existing models and available data do not adequately capture transmission complexity, reporting gaps, population density, lockdowns, variants, travel, and social factors.Some relevant factors, including culture and lifestyle, are qualitative and cannot be measured straightforwardly.
- 1 Introduction: Recurrent neural networks suit spatiotemporal sequences, while LSTMs address the long-term dependency limitations of standard RNN training.Prior work had applied LSTMs to COVID-19 forecasting in China and Canada.
- 1 Introduction: The study applies LSTM models to selected Indian COVID-19 hotspots and provides short-term forecasts using data from both infection waves.The paper compares univariate and multivariate approaches for four-days-ahead forecasting and provides two-month-ahead forecasts.
2 Related Work
Prior COVID-19 forecasting studies used statistical, machine-learning, ecological, and deep-learning models across countries and Indian states, with mixed predictive performance. Indian work increasingly examined recurrent networks, environmental variables, preventive measures, and longer-horizon case forecasting.
- International approaches: Researchers applied autoregressive, ARIMA, Gaussian-process, ecological-niche, and LSTM models to forecast infections, peaks, transmission, and geographic risk zones.These studies covered Egypt, Pakistan, the United States, Canada, China, and megacities including Beijing, Guangzhou, and Shenzhen.
- Forecasting performance: COVID-19 forecasting studies reported mixed accuracy, including substantial underprediction of cases in Pakistan and the United States despite closer death estimates in the latter.The ARIMA model predicted 5,681–33,079 Pakistani cases by late May 2020 versus approximately 70,000 reported cases; the U.S. Gaussian-process model also substantially underpredicted peak infections.
- Deep-learning approaches: LSTM-based forecasting in Canada was relatively close to the observed timing of the daily new-case peak, although it focused on a single national outbreak trajectory.The study predicted a possible outbreak endpoint around June 2020, while Canada reached its daily new-case peak on 2 May 2020.
- Indian forecasting studies: Indian studies used LSTM, GRU, weather-informed models, preventive-measure analysis, and 30-day-ahead prediction to model cases across states.The reviewed work included data from January–July 2020, state-level weather conditions, transmission-rate assumptions, and preventive interventions.
- Hybrid and comparative models: Broader forecasting research also combined machine learning with deep learning or Bayesian models to estimate COVID-19 trajectories across multiple countries.One cited logistic-growth model estimated late-September 2020 totals for the United States, Brazil, India, and Russia.
3 Methodology: Forecasting COVID-19 novel infections with deep learning models
The methodology reconstructs infection time series for multi-step forecasting and applies LSTM variants that use temporal memory, bidirectional context, or sequence-to-sequence mapping. It also characterizes Indian state-level infection patterns across the 2020 and 2021 waves using reported cases and visual summaries.
- 3 Methodology: Forecasting COVID-19 novel infections with deep learning models: Takens’ theorem reconstructs an embedded phase-space vector from delayed observations for training multi-step-ahead deep-learning models.The embedding uses time delay T, embedding dimension D, and a window of lagged observations.
- 3.1 LSTM network models: LSTM models address simple RNN difficulties with long-term dependencies by using memory cells in the hidden layer.The methodology describes vanishing and exploding gradients in simple RNNs and presents LSTM memory cells as the response.
- 3.2 Bi-directional LSTM networks: Bidirectional LSTM networks combine forward and backward sequence processing so each output can preserve information from both past and future states.Two hidden-state sequences are combined at each time point.
- 3.3 Encoder-Decoder LSTM networks: Encoder-decoder LSTMs map variable-length input sequences to variable-length outputs through encoding hidden states and decoding an output distribution.This sequence-to-sequence structure supports multi-step series prediction when input and output lengths differ.
- 3.4 India: Situation Report: 8th October, 2021: The situation analysis ranks states by monthly novel totals and visualizes weekly infections, active cases, and cumulative deaths across Indian regions and waves.The comparisons highlight differences among populous states, infection peaks, and timing of state-level peaks.
4 Results
The study evaluates LSTM, BD-LSTM, and ED-LSTM models for univariate and multivariate COVID-19 forecasting in India, using multiple data-splitting strategies and prediction horizons. ED-LSTM generally performs best for univariate models, whereas multivariate results favor ED-LSTM with static splitting, and recursive forecasts show declining cases with short-term fluctuations.
- Experimental Design: The experiments use normalized daily case data, three-day rolling means, Taken’s-theorem state-space reconstruction, and four multi-step prediction horizons.The reconstruction uses embedding dimension D = 6 and time-lag T = 2, with MSA = 4 defining the prediction horizons.
- Experimental Design: Multivariate inputs represent four adjacent states for each target state, enabling models to incorporate neighboring-state case information.Examples include Maharashtra with Gujarat, Madhya Pradesh, and Uttar Pradesh, and Delhi with Rajasthan, Uttar Pradesh, and Haryana.
- Prediction performance: ED-LSTM provides the best performance for univariate models under random-split evaluation, with lower RMSE than the other tested models in general.The univariate comparison includes static and random train/test splits across India, Maharashtra, and Delhi.
- Prediction performance: ED-LSTM with static-split provides the best overall multivariate performance, while random-split multivariate results sometimes favor BD-LSTM.The comparison covers LSTM, BD-LSTM, and ED-LSTM across test datasets and prediction horizons.
- Prediction performance: Recursive univariate forecasts for October and November 2021 show generally declining cases, while LSTM models capture recurring short-term spikes and falls.Forecast uncertainty is reported using 95% confidence intervals over 30 experiment runs.
5 Discussion
The discussion emphasizes that forecasting uncertainty reflects complex infection dynamics, limited data, and unmeasured social and demographic factors. Forecasts indicate declining cases, but improved uncertainty quantification and richer features are needed for more reliable long-term predictions.
- India’s pandemic included two major peaks, with the second wave associated with a highly infectious Delta variant and inadequate preparation and resources.The discussion links the second wave to multiple factors, including shortages of oxygen and temporary hospitals and poor lockdown management.
- The two-month forecasts showed a general decline in new cases, but uncertainty differed across India, Delhi, and Maharashtra because their historical case trends contained different numbers of peaks.Delhi and Maharashtra had multiple peaks or outbreaks, making their predictions less certain than those for India overall.
- Forecasting is constrained by missing information on social-cultural factors, population density, lockdown levels, travel behavior, restrictions, vaccination, and access to information.The authors state that incorporating these features could improve forecasting results.
- Population density, migration, rural residence, cultural rituals, and family structure make cross-country or cross-city comparisons difficult for computational models.The authors caution that similar overall populations do not imply comparable population dynamics or transmission conditions.
- Future work should use robust uncertainty quantification, including Bayesian deep learning, and could combine LSTM variants through ensemble learning.The authors also propose forecasting deaths and jointly modelling case trends and economic effects.
6 Conclusions
The paper presents an LSTM-based framework for forecasting daily COVID-19 infections in India and evaluates multiple model and data configurations. The results show a general decline in projected cases, while limited and biased data make model selection and long-term forecasting difficult.
- The framework applies prominent deep learning forecasting tools to daily novel infection prediction in India while highlighting challenges caused by limited data and complex infection dynamics.
- The univariate random-split ED-LSTM model provided the best test performance among the evaluated models, while multivariate models did not outperform it.The authors found that adjacent-state data had limited effect in the multivariate setting.
- The two-month forecast showed a general decline in new cases, but authorities were advised to remain vigilant.The conclusion does not identify a single universally superior LSTM variant across all scenarios.