📊 statistics

Dynamic Prediction of Alternating Recurrent Events via Neural Network

This paper introduces a novel online dynamic prediction framework that combines neural network theory with inverse probability weighted pseudo-observations to accurately forecast alternating recurrent events, demonstrating strong performance in simulations and successfully predicting low-mood periods among first-year medical residents.

Original authors: Abigail Loe, Susan Murry, Zhenke Wu

Published 2026-07-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Abigail Loe, Susan Murry, Zhenke Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the "On" and "Off" Switches of Life

Imagine your life (or your health) isn't a straight line, but a series of switches. Sometimes you are "on" (like being happy), and sometimes you are "off" (like being in a low mood). When you flip to "off," you can't immediately flip back to "on." You have to wait for a "recovery period" before you are allowed to flip the switch again.

In statistics, these are called Alternating Recurrent Events.

The Event: A period of low mood.
The Refractory Period: The time you spend recovering, during which you cannot have a new low mood episode because you are already in one.

The problem is that predicting when the next "low mood" will start is tricky. If you are currently in a low mood, you aren't at risk for a new one yet. If you are happy, you are at risk. Traditional math tools often get confused by this "switching" behavior and the fact that people drop out of studies (censoring) before we see the whole picture.

The Solution: A "Smart" Digital Assistant

The authors of this paper built a new tool to predict these switches in real-time. They combined two powerful ideas:

The "Pseudo-Observation" Trick:
Imagine you are trying to guess how long a movie will last, but some people leave the theater early. Instead of just ignoring them, the authors use a statistical "magic trick" (called pseudo-observations) to estimate what those early leavers would have seen if they stayed. This fills in the gaps so the prediction isn't biased.
The "Memory" Machine (LSTM):
They used a type of Artificial Intelligence called a Long Short-Term Memory (LSTM) network. Think of this as a digital assistant with an excellent memory.
- Old tools might look at your mood today and guess your mood tomorrow.
- This LSTM looks at your mood today, remembers your mood from last week, last month, and even the pattern of your previous "low" and "high" periods. It learns the rhythm of your life to make a better guess.

How They Tested It: The "Video Game" Simulation

Before trying this on real people, the authors built a virtual world (a simulation) to test their tool.

They created 100 fake people with complex, made-up lives.
They programmed these people to have "low moods" and "recovery periods" based on specific rules.
They tested their new AI tool against older, standard math tools.

The Results:

The Old Tools: Were okay, but they struggled when the data got messy or when people had a lot of history to remember.
The New AI Tool: As the simulation went on and the AI got to know the "people" better (more data), it became incredibly accurate. It was better at spotting the patterns and predicting exactly when the next "low mood" would start.

The Real-World Test: Medical Residents

The authors then applied their tool to real data from the Intern Health Study.

The Subjects: First-year medical residents (doctors just out of med school).
The Goal: Predict when a resident's mood would drop significantly below their normal baseline.
The Data: The residents reported their mood daily, along with data like how many steps they took, how they slept, and their stress levels.

What Happened:
The AI tool was outstanding. It predicted mood drops with much higher accuracy than the standard statistical methods.

It correctly identified the "at-risk" moments about 97% of the time (a very high score in this field).
It was able to learn from the residents' past history (like "I usually get low after a week of bad sleep") to make better predictions for the future.

The Catch: The "Black Box"

The paper admits one major limitation.

Old Math Tools: You can look at the result and say, "Ah, the math shows that lack of sleep increases the risk by 10%." It's transparent.
The AI Tool: It works like a "black box." It gives you a very accurate prediction, but it's hard to explain exactly how it decided that. It found complex patterns that human math formulas can't easily describe.

Summary

The paper presents a new way to predict recurring life events (like mood swings or hospital stays) that alternate with recovery periods. By using a "memory-rich" AI that learns from a person's entire history and correcting for missing data, they created a tool that predicts future events much better than traditional methods. They proved this works in computer simulations and on real medical residents, though the AI remains a bit mysterious about how it reaches its conclusions.

Technical Summary: Dynamic Prediction of Alternating Recurrent Events via Neural Network

Problem Statement
Alternating recurrent events—outcomes where a primary event (e.g., hospitalization, depressive episode) triggers a secondary refractory period during which the individual is not at risk for the primary event—occur frequently in biostatistics, behavioral science, and criminal justice. Analyzing these events presents unique statistical challenges: the two event types are mutually exclusive, creating competing risks; observations are correlated; and data are subject to censoring. Traditional approaches often rely on state-based Markov processes or semi-parametric intensity models (e.g., Cox-based extensions). However, these methods often struggle with the "dependent gap-time" issue or require analysts to specify parametric relationships between covariates and outcomes. Furthermore, existing dynamic prediction tools for recurrent events rarely address "online" or real-time prediction, where an individual's risk profile is updated continuously based on newly observed experience, particularly for alternating event structures.

Methodology
The authors propose an online dynamic prediction framework that combines pseudo-observations with deep learning to predict the time to the next primary alternating recurrent event over a fixed follow-up window ( $\tau$ ). The methodology proceeds through four key stages:

Data Transformation: Following the approach of Tayob and Murray (2015), the authors transform the alternating recurrent event timeline into a censored longitudinal dataset. At prespecified check-in times $t$ , the data is restructured to identify the time-to-first-observed event after $t$ . Crucially, this structure accounts for the "at-risk" status ( $R_i(t)$ ); subjects currently in a refractory (secondary) period are marked as not at-risk for the primary event, rendering their primary event time unobserved at that moment.
Pseudo-Observations: To handle independent censoring, the authors construct pseudo-observations for the $\tau$ -restricted mean survival time, $E[\min\{T_i(t), \tau\}]$ . Using a jack-knife methodology (Andersen et al., 2003), these pseudo-observations replace the censored survival outcome, allowing for regression-based modeling.
Inverse Probability Weighting (IPW): To correct for bias introduced by the non-random selection of subjects into the "at-risk" set (due to the alternating nature of the events), the framework employs inverse probability weights. The probability of being at-risk, $P(R_i(t)=1 | \bar{Z}_i(t))$ , is estimated using either a Long Short-Term Memory (LSTM) network or a longitudinal Random Forest. These weights are incorporated into the loss function of the prediction model.
Prediction Models: Two dynamic prediction models are proposed:
- LSTM-based Model: A neural network architecture utilizing LSTMs to model the pseudo-observations. This approach leverages gating mechanisms to incorporate historical covariate data and handles complex, non-linear relationships without pre-specifying functional forms.
- Weighted Semi-parametric Model: A Generalized Estimating Equation (GEE) approach that fits a linear model to the weighted pseudo-observations. This serves as a semi-parametric benchmark offering interpretability.

Key Contributions

Novel Framework: This is the first proposed dynamic prediction model specifically designed for alternating recurrent events in an online setting.
Integration of ML and Statistics: The paper bridges the gap between statistical survival analysis and modern machine learning by adapting LSTM architectures for censored longitudinal data using pseudo-observations and IPW.
Handling of Refractory Periods: The method explicitly addresses the "at-risk" status, ensuring that predictions for the primary event are not biased by time spent in the secondary (refractory) state.
Comparison of Weighting Strategies: The study evaluates whether LSTM or Random Forest is superior for estimating the inverse probability weights required to debias the prediction models.

Results
The proposed methods were evaluated through simulation studies and an application to the Intern Health Study (IHS) data, predicting periods of low mood for first-year medical residents.

Simulation: In scenarios with complex covariate relationships and correlated gap times, the LSTM-based dynamic prediction model demonstrated superior performance compared to semi-parametric GEE models (both "Oracle" and "Main Effects") in terms of the C-statistic (discrimination) and Root Error Sum of Squares (RESS), particularly as historical data accumulated. While the LSTM showed slightly higher bias in some settings (a classic bias-variance trade-off), its predictive discrimination was outstanding, often exceeding the semi-parametric oracle model. Both LSTM and Random Forest performed similarly when used to estimate the inverse probability weights.
Application (IHS): When applied to real-world data, the LSTM model achieved a maximum C-statistic of approximately 0.97, significantly outperforming the GEE Main Effects model (max C-statistic ~0.82). The LSTM model using LSTM-derived weights yielded the lowest RESS after the first 300 days of follow-up.

Significance and Claims
The authors claim that their framework provides a robust tool for real-time clinical decision-making, specifically for identifying opportune moments for intervention (e.g., mobile app prompts for medical residents). The significance of the work lies in its ability to leverage the predictive power of deep learning (specifically LSTMs) while maintaining statistical rigor through pseudo-observations and inverse probability weighting to handle censoring and alternating event structures.

The paper modestly acknowledges that while the LSTM approach offers superior predictive accuracy, it operates as a "black box" compared to the semi-parametric GEE approach. The lack of interpretable parameter estimates in the neural network model means the specific functional form of how history covariates influence the outcome remains opaque, despite the use of embeddings. The authors emphasize that the success of the LSTM relies heavily on careful hyperparameter selection and the availability of sufficient longitudinal history, noting that the method is not robust to random hyperparameter choices. They conclude that while machine learning methods excel in prediction, the field still lags in providing the theoretical statistical proofs (consistency, identifiability) that underpin traditional statistical methods.