Chapter 6: Predictive Analytics & Crisis Detection

6.1 Real-Time Monitoring Systems

Traditional mental healthcare relies on periodic check-ins (weekly therapy sessions, monthly medication reviews). Digital therapeutics enable continuous monitoring, detecting concerning patterns immediately rather than waiting for the next appointment.

Monitoring Data Streams

Real-Time Monitoring Architecture

class RealTimeMonitoringEngine {
  private streamProcessor: StreamProcessor;
  private riskModels: Map;
  private alertDispatcher: AlertDispatcher;
  
  async processDataStream(
    userId: string,
    dataPoint: DataPoint
  ): Promise {
    // 1. Validate and normalize data
    const normalized = await this.normalize(dataPoint);
    
    // 2. Update user's time series
    await this.updateTimeSeries(userId, normalized);
    
    // 3. Get recent history for context
    const history = await this.getRecentHistory(userId, {
      lookback: 14 // days
    });
    
    // 4. Calculate current risk scores
    const risks = await this.assessRisks(userId, normalized, history);
    
    // 5. Check for alert thresholds
    const alerts = this.checkThresholds(risks);
    
    // 6. Dispatch alerts if needed
    if (alerts.length > 0) {
      await this.dispatchAlerts(userId, alerts);
    }
    
    // 7. Update dashboards
    await this.updateDashboards(userId, risks);
  }
  
  private async assessRisks(
    userId: string,
    current: DataPoint,
    history: TimeSeriesData
  ): Promise {
    // Parallel risk assessment across multiple models
    const [
      crisisRisk,
      relapseRisk,
      dropoutRisk,
      deteriorationRisk
    ] = await Promise.all([
      this.riskModels.get('crisis').predict(userId, current, history),
      this.riskModels.get('relapse').predict(userId, current, history),
      this.riskModels.get('dropout').predict(userId, current, history),
      this.riskModels.get('deterioration').predict(userId, current, history)
    ]);
    
    return {
      crisis: {
        probability: crisisRisk.probability,
        confidence: crisisRisk.confidence,
        timeframe: '24-72 hours',
        factors: crisisRisk.contributingFactors
      },
      relapse: {
        probability: relapseRisk.probability,
        confidence: relapseRisk.confidence,
        timeframe: '1-4 weeks',
        factors: relapseRisk.contributingFactors
      },
      dropout: {
        probability: dropoutRisk.probability,
        confidence: dropoutRisk.confidence,
        timeframe: '1-2 weeks',
        factors: dropoutRisk.contributingFactors
      },
      deterioration: {
        probability: deteriorationRisk.probability,
        confidence: deteriorationRisk.confidence,
        timeframe: '1 week',
        factors: deteriorationRisk.contributingFactors
      },
      overallRiskLevel: this.calculateOverallRisk([
        crisisRisk,
        relapseRisk,
        dropoutRisk,
        deteriorationRisk
      ])
    };
  }
}

6.2 Crisis Prediction Models

The most critical application of predictive analytics is identifying patients at imminent risk of suicide or self-harm. ML models trained on patterns preceding past crises can provide early warning, enabling life-saving interventions.

Crisis Risk Factors

Crisis Prediction Implementation

class CrisisPredictionModel {
  private lstm: LSTMModel;  // For time series patterns
  private bert: BERTModel;  // For text analysis
  private ensemble: EnsembleModel;
  
  async predictCrisisRisk(
    userId: string,
    timeframe: '24h' | '48h' | '72h'
  ): Promise {
    // Gather multi-modal data
    const data = await this.gatherData(userId);
    
    // Time series analysis (14-day patterns)
    const timeSeriesFeatures = await this.analyzeTimeSeries({
      mood: data.moodRatings,
      sleep: data.sleepData,
      activity: data.activityData,
      engagement: data.platformEngagement
    });
    
    // Text analysis (recent journal entries, messages)
    const textFeatures = await this.analyzeText({
      journalEntries: data.recentJournals,
      messages: data.recentMessages,
      thoughtRecords: data.thoughtRecords
    });
    
    // Clinical assessments
    const clinicalFeatures = {
      phq9Trend: this.calculateTrend(data.phq9Scores, 30),
      gad7Trend: this.calculateTrend(data.gad7Scores, 30),
      currentSeverity: data.latestPHQ9,
      previousCrises: data.crisisHistory.length,
      riskFactors: data.demographicRiskFactors
    };
    
    // Combine features
    const combinedFeatures = {
      ...timeSeriesFeatures,
      ...textFeatures,
      ...clinicalFeatures
    };
    
    // Ensemble prediction
    const prediction = await this.ensemble.predict(combinedFeatures);
    
    // Risk stratification
    let riskLevel: 'low' | 'moderate' | 'high' | 'severe';
    if (prediction.probability >= 0.7) riskLevel = 'severe';
    else if (prediction.probability >= 0.5) riskLevel = 'high';
    else if (prediction.probability >= 0.3) riskLevel = 'moderate';
    else riskLevel = 'low';
    
    return {
      riskLevel,
      probability: prediction.probability,
      confidence: prediction.confidence,
      timeframe,
      contributingFactors: this.explainPrediction(
        combinedFeatures,
        prediction
      ),
      recommendedActions: this.getActions(riskLevel),
      requiresImmediateAction: riskLevel === 'severe' || riskLevel === 'high',
      predictedAt: new Date()
    };
  }
  
  private getActions(riskLevel: string): Action[] {
    const actions = {
      severe: [
        {
          action: 'NOTIFY_EMERGENCY_CONTACTS',
          priority: 'immediate',
          description: 'Notify designated emergency contacts'
        },
        {
          action: 'ALERT_THERAPIST',
          priority: 'immediate',
          description: 'Urgent notification to assigned therapist'
        },
        {
          action: 'DISPLAY_CRISIS_RESOURCES',
          priority: 'immediate',
          description: '988 Suicide & Crisis Lifeline prominent display'
        },
        {
          action: 'ACTIVATE_SAFETY_PLAN',
          priority: 'immediate',
          description: 'Guide through personalized safety plan'
        },
        {
          action: 'SCHEDULE_URGENT_SESSION',
          priority: 'high',
          description: 'Offer same-day or next-day session'
        }
      ],
      high: [
        {
          action: 'ALERT_THERAPIST',
          priority: 'high',
          description: 'Notify therapist for follow-up within 24h'
        },
        {
          action: 'OFFER_IMMEDIATE_SUPPORT',
          priority: 'high',
          description: 'Suggest crisis chat or helpline'
        },
        {
          action: 'REVIEW_SAFETY_PLAN',
          priority: 'medium',
          description: 'Prompt to review and update safety plan'
        }
      ],
      moderate: [
        {
          action: 'THERAPIST_NOTIFICATION',
          priority: 'medium',
          description: 'Inform therapist for monitoring'
        },
        {
          action: 'INCREASE_MONITORING',
          priority: 'medium',
          description: 'Suggest daily check-ins'
        },
        {
          action: 'COPING_SKILLS_REMINDER',
          priority: 'low',
          description: 'Remind of learned coping strategies'
        }
      ],
      low: [
        {
          action: 'ROUTINE_MONITORING',
          priority: 'low',
          description: 'Continue standard monitoring'
        }
      ]
    };
    
    return actions[riskLevel] || actions.low;
  }
}

6.3 Automated Intervention Systems

When risks are identified, automated systems can deliver appropriate interventions immediately rather than waiting for human clinicians to review alerts. These "just-in-time adaptive interventions" (JITAIs) provide support precisely when needed.

JITAI Framework

class JITAIEngine {
  private interventionLibrary: InterventionRepository;
  private userPreferences: Map;
  private effectivenessTracker: EffectivenessTracker;
  
  async selectIntervention(
    userId: string,
    context: Context,
    risk: RiskAssessment
  ): Promise {
    // Get user's intervention history and preferences
    const history = await this.getUserHistory(userId);
    const preferences = this.userPreferences.get(userId);
    
    // Get candidate interventions for this risk level
    const candidates = await this.interventionLibrary.search({
      riskLevel: risk.level,
      riskType: risk.type,
      context: context.situation
    });
    
    // Score each intervention
    const scored = candidates.map(intervention => {
      let score = 0;
      
      // Base effectiveness for this risk type
      score += this.effectivenessTracker.getEffectiveness(
        intervention.id,
        risk.type
      ) * 40;
      
      // Personal effectiveness history
      const personalHistory = history.filter(
        h => h.interventionId === intervention.id
      );
      if (personalHistory.length > 0) {
        const avgRating = personalHistory.reduce(
          (sum, h) => sum + h.helpfulnessRating, 
          0
        ) / personalHistory.length;
        score += avgRating * 30;
      }
      
      // User preferences
      if (preferences.preferredTypes.includes(intervention.type)) {
        score += 20;
      }
      
      // Contextual appropriateness
      if (this.isContextuallyAppropriate(intervention, context)) {
        score += 10;
      }
      
      // Novelty (avoid overuse)
      const recentUse = history.filter(
        h => h.interventionId === intervention.id &&
        h.timestamp > Date.now() - 7 * 24 * 60 * 60 * 1000
      ).length;
      score -= recentUse * 5;
      
      return { intervention, score };
    });
    
    // Select highest-scoring intervention
    const selected = scored.sort((a, b) => b.score - a.score)[0];
    
    // Log selection for learning
    await this.logSelection({
      userId,
      intervention: selected.intervention,
      context,
      risk,
      score: selected.score
    });
    
    return selected.intervention;
  }
  
  async deliverIntervention(
    userId: string,
    intervention: Intervention,
    context: Context
  ): Promise {
    // Choose delivery channel based on urgency and context
    const channel = this.selectChannel(intervention.urgency, context);
    
    // Personalize intervention content
    const personalized = await this.personalizeContent(
      intervention,
      userId
    );
    
    // Deliver via selected channel
    const result = await channel.send({
      userId,
      title: personalized.title,
      content: personalized.content,
      actions: personalized.actions,
      priority: intervention.urgency
    });
    
    // Track delivery and response
    await this.trackDelivery({
      userId,
      interventionId: intervention.id,
      deliveredAt: new Date(),
      channel: channel.name,
      result
    });
    
    // Set up response monitoring
    this.monitorResponse(userId, intervention, result);
    
    return result;
  }
}

6.4 Relapse Prevention

For conditions like depression with high relapse rates (50-80% lifetime risk), early detection of warning signs enables timely intervention before full relapse occurs.

Relapse Signatures