💻 computer science

Predicting Case Processing Duration in Kenyan Magistrate Courts: A Three-Stage Machine Learning Pipeline for Proactive Backlog Management

This study introduces a three-stage machine learning pipeline, validated on over 670,000 Kenyan magistrate court cases, which utilizes LightGBM and CatBoost models to predict case processing durations and classify same-day disposals, thereby enabling proactive backlog management through an empirically derived risk scheme and a deployed web application.

Original authors: Anuary Mulombi, Fidelis Mukudi, Anthony Mile

Published 2026-08-10

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Anuary Mulombi, Fidelis Mukudi, Anthony Mile

Original paper licensed under CC BY 4.0 (https://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

Imagine a giant, bustling library where millions of books are constantly being checked in and out. In this library, the librarians are overwhelmed. Some books are returned the same day they are checked in, but others get lost in the stacks for years, creating a massive pile of "backlog" that no one can seem to clear. This isn't just about books; it's about justice. When cases (the "books") sit too long, people lose faith in the system, and their rights to a fair and speedy trial are delayed. For a long time, librarians could only count the books they had already processed; they had no way to guess which new books would get stuck in the pile before they even left the desk.

Enter the world of Machine Learning, a branch of computer science where we teach computers to find patterns in history, much like a detective studying old crime scenes to solve new mysteries. Instead of guessing, these computers look at thousands of past cases to learn what makes a case move fast or get stuck. The key idea here is predictive analytics: using data from the past to forecast the future. If a computer can tell us, "This new case looks like the ones that took five years," we can stop it from getting lost in the pile before it even starts. This is exactly what the researchers in this study set out to do for the courts in Kenya, turning a reactive system (fixing problems after they happen) into a proactive one (stopping problems before they grow).

The Three-Stage Crystal Ball for Court Cases

In the Kenyan magistrate courts, where more than 80% of all legal business happens, a massive backlog has been building up. As of mid-2025, there were over 127,000 cases waiting to be heard. The existing computer system, called the Integrated Case Management System (ICMS), is great at recording what has happened, but it's blind to what might happen. It doesn't know which new cases are likely to become the "super-slow" ones that clog the system.

To fix this, the researchers built a three-stage machine learning pipeline. Think of this pipeline as a high-tech sorting machine at a post office that doesn't just stamp letters; it predicts how long each letter will take to reach its destination and flags the ones that might get lost.

Stage 0: The "Same-Day" Speed Trap

The first job of this machine is to spot the "express" cases. About 29% of all cases in Kenya are resolved on the very day they are filed. The researchers trained a computer model to look at a new case the moment it arrives and ask: "Is this a quick one, or is it going to be a marathon?"
They tested four different computer "brains" (algorithms) to see which was best at this. The winner was LightGBM, which correctly identified same-day cases with a high level of accuracy (a score of 0.9148). If the computer says "Same Day," the case gets a green light and moves on without further delay. If it says "Multi-Day," the case moves to the next stage.

Stage 1: The "Filing-Time" Crystal Ball

Once a case is flagged as a "multi-day" marathon, the system needs to guess how long the race will take. This is the tricky part because the case hasn't even started its journey yet. The computer looks only at the information available the moment the case is filed: the type of case, how many lawyers are involved, how busy the specific judge is, and whether the court is in a city or a rural area.
Using data from nearly 673,000 past cases, the LightGBM model again proved to be the champion. It could predict the duration of a case with a margin of error of about 444 days. While that sounds like a lot, it explained about 62% of the reasons why cases take as long as they do. This means that just by looking at the paperwork on day one, the system can tell a judge, "This case looks like it will take about 1.5 years," or "This one might drag on for four years."

Stage 2: The "Live-Update" Tracker

The journey doesn't end at the filing desk. As the case moves through the court, it accumulates history: hearings are held, adjournments are granted, and delays happen. The third stage of the pipeline is a dynamic model that updates the prediction as the case progresses.
Imagine a GPS that doesn't just give you an arrival time when you leave home, but recalculates it every time you hit a traffic jam. This Stage 2 model (powered by a different computer brain called CatBoost) takes the new information—like how many hearings have happened or how long the first meeting was delayed—and revises the estimate.
The results here were impressive. While the initial guess (Stage 1) had an error of 444 days, the updated guess (Stage 2) narrowed that error down to just 80.83 days. This shows that as the case moves forward, the computer gets much better at knowing exactly when it will finish.

The "JaliHaki" Dashboard: Putting It All to Work

The researchers didn't just stop at the math; they built a real-world tool called JaliHaki (which means "to guard or protect justice" in Swahili). This is a web application that connects directly to the court's computer system.

For Court Administrators: It shows a dashboard with a list of cases, color-coded by risk. A "Low Risk" case is green and needs routine monitoring. A "Critical Risk" case is red and needs immediate attention from a senior judge.
For Magistrates: It gives them a personal view of their workload, warning them if a case assigned to them is starting to drift into the "Critical" zone.
For Litigants (The Public): It offers a plain-language estimate. Instead of confusing legal jargon, a person filing a case might see: "Your case is estimated to take 5 years, and it is currently in the High Risk category."

What Drives the Delays?

By using a technique called SHAP analysis (which acts like a magnifying glass to see which factors matter most), the researchers discovered the main culprits behind delays:

Caseload per Magistrate: When a judge has too many cases, everything slows down.
Legal Representation: Whether a party has a lawyer or not changes the timeline.
Case Nature: Criminal cases tend to move faster (median 90 days for multi-day cases) than civil cases (median 547 days).
First-Mention Delay: How long it takes to hold the very first hearing is a huge predictor of the final delay.
Accumulated Hearings: The more hearings a case has, the longer it tends to take.

Interestingly, the study found that Urban courts (in big cities) actually took longer to process multi-day cases (median 275 days) than Rural courts (median 99 days). This seems counter-intuitive, but the researchers explain it using a concept called Queueing Theory: even if a city court is well-resourced, the sheer volume of cases creates a longer "waiting line," slowing everyone down.

The Four-Tier Risk System

To make the predictions useful for real people, the team created a simple four-tier risk system based on the computer's predictions:

Low Risk (≤ 218 days): Routine monitoring.
Medium Risk (218–564 days): Scheduled review.
High Risk (564–1,283 days): Immediate intervention needed.
Critical Risk (> 1,283 days): Escalate to a senior magistrate immediately.

This system allows the court to spot the "Critical Risk" cases (which make up about 10.5% of new filings) right at the start and take action before they become part of the massive backlog.

What the Study Does and Doesn't Say

The researchers are careful to note that their model is a tool for prediction, not a magic wand. They explicitly state that the model learns from historical data, which means it reflects the current realities of the court system, including any existing inequalities between different counties or court types. They warn that the model's accuracy for rural courts might need more testing because there were fewer rural cases in their data.

They also clarify that the "filing year" was a top predictor, but not because the courts are getting faster or slower over time. Instead, it's a mathematical quirk: cases filed recently haven't had enough time to finish yet, so the computer sees them as "shorter" simply because they are new. The researchers adjusted for this to ensure the predictions were fair.

The Bottom Line

This study proves that with the data courts already have, we can build a system that predicts case delays with surprising accuracy. By using a three-stage pipeline, the Kenyan courts can now move from reacting to backlogs to preventing them. The JaliHaki application turns complex math into a simple, color-coded dashboard that helps judges, clerks, and the public understand the timeline of justice. While the model isn't perfect (it still has an error margin of about 80 days for updated predictions), it represents a massive leap forward in using technology to protect the right to a speedy trial. The study suggests that if these tools are adopted, courts can identify the cases that need help the most, right from the moment they are filed, ensuring that justice doesn't get lost in the stacks.

Technical Summary: Predicting Case Processing Duration in Kenyan Magistrate Courts

Problem Statement
Kenyan magistrate courts handle over 80% of the nation's judicial business but suffer from a severe case backlog crisis, accounting for 75% of the national backlog as of June 2025. While the Judiciary's Integrated Case Management System (ICMS) records historical case events, it lacks predictive capabilities to identify which newly filed cases are likely to experience delays at the point of filing. Consequently, case management remains reactive, identifying backlogs only after they have materialized, which infringes upon the constitutional right to justice without delay.

Methodology
The study utilizes a census extraction of 744,340 closed cases filed between January 2011 and March 2026. After cleaning and validation, 672,974 records were retained. The research proposes and validates a three-stage machine learning pipeline:

Stage 0 (Triage): A binary classifier distinguishing between "same-day disposals" (28.98% of cases) and "multi-day proceedings." This stage uses a stratified 80/20 split on the full dataset.
Stage 1 (Filing-Time Prediction): A regression model predicting total case duration using only features available at the time of filing (e.g., case nature, legal representation, magistrate caseload). This stage is trained exclusively on the 477,952 multi-day cases using a random 80/20 split.
Stage 2 (Dynamic Progression): A regression model that updates the duration prediction as procedural history accumulates (e.g., number of hearings, adjournment delays). This stage uses a time-based 80/20 split (earliest 80% for training, most recent 20% for testing) to simulate real-world deployment.

Data and Features
The study introduces a formal classification of court stations into Urban, Semi-Urban, and Rural categories based on the Kenya Urban Areas and Cities Act and Census data. Feature engineering includes:

Group A (Filing-time): Temporal features, staffing ratios, case complexity proxies, and court specialization.
Group B (Dynamic): Procedural history such as first-mention delay, total adjournments, and hearing efficiency.
Exclusion: The final case outcome was explicitly excluded to prevent data leakage.

Algorithm Selection and Tuning
Four ensemble algorithms were evaluated: Random Forest, XGBoost, LightGBM, and CatBoost. Hyperparameter tuning involved baseline fitting, RandomizedSearchCV, and GridSearchCV under grouped five-fold cross-validation (grouped by court_id) to prevent information leakage. SHAP (SHapley Additive exPlanations) was used for model interpretability.

Key Results

Stage 0 (Classification): LightGBM achieved the best performance with a ROC-AUC of 0.9148 and a recall of 0.8659, effectively triaging same-day cases.
Stage 1 (Filing-Time Regression): LightGBM performed best with an RMSE of 444.25 days and an $R^2$ of 0.6157, explaining approximately 62% of the duration variance using only filing-time data.
Stage 2 (Dynamic Regression): CatBoost achieved the lowest RMSE of 80.83 days on the chronologically held-out test set. While the $R^2$ was lower (0.2724) due to right-censoring in the recent test set, the absolute error reduction (from ~444 to ~81 days) demonstrates the value of procedural history.
Feature Importance: SHAP analysis identified caseload per magistrate, legal representation, case nature, first-mention delay, and accumulated hearings as dominant drivers of delay.
Risk Classification: An empirically derived four-tier risk scheme (Low, Medium, High, Critical) was established using percentiles of the predicted duration distribution, with thresholds at 218, 564, and 1,283 days.

Implementation: JaliHaki Web Application
The models were operationalized into JaliHaki, a Laravel-based web application integrated with a Flask REST API. The system provides:

Court Administrator Dashboard: Displays caseload by risk tier, compliance metrics, and allows "Update Prediction" actions to trigger Stage 2 models.
Magistrate Dashboard: Shows active caseload summaries and risk classifications for assigned cases.
Litigant Portal: Provides plain-language duration estimates and risk badges upon filing.
The workflow is atomic: upon case submission, Stage 0 and Stage 1 are invoked; if the case is multi-day, a risk tier is assigned. Stage 2 is triggered whenever new proceedings are recorded.

Significance and Claims
The paper claims three primary contributions:

Methodological Innovation: It is the first study to combine a same-day disposal classifier, a filing-time regression model, and a dynamic progression regression model into a single pipeline for any judicial system.
Contextual Application: It applies this architecture specifically to Kenyan magistrate court data, covering both civil and criminal cases, extending previous work that was often limited to criminal cases or descriptive analysis.
Operational Integration: Unlike prior studies that remain theoretical, this work produces a deployable web application (JaliHaki) that integrates the pipeline directly into judicial administration workflows.

The authors assert that meaningful, evidence-based case duration prediction is achievable using existing administrative records without new data infrastructure. The study concludes that shifting from reactive to proactive backlog management is feasible, allowing for immediate identification of "Critical Risk" cases (approx. 10.5% of multi-day cases) that require escalation to senior magistrates.

Limitations
The authors note three constraints:

Right Censoring: Recently filed cases have less time to close, affecting Stage 2 metrics and the prominence of the filing_year feature.
Rural Under-representation: While 34% of stations are rural, they contribute only 2.3% of case records, necessitating further validation for rural-specific performance.
Search Tractability: Hyperparameter searches were conducted on subsamples (up to 40,000 records), suggesting potential for marginal improvements with more exhaustive searches.